Skip to main content
Ask Hive is the fastest way to get an answer only your own systems can give. It reads across your connected tools, uploaded files, the Business Brain and — when a question needs it — the web, then answers with citations you can open and check. Open it with New chat in the sidebar. Hive opens on Ask too.
Ask is part of every Basic, Pro and Max seat. A member on an Approver seat can ask Hive only about the items routed to them, from the approval card or Inbox item in the Hive app, 10 questions a calendar month. See The Approver seat.
This page covers Ask from end to end. If you came for one thing, start here:

Ask a question

1

Type your question

Use the composer: “Ask anything — or pick up where an agent left off…”. Questions can be up to 20,000 characters. A long paste — more than 2,000 characters or 40 lines — is attached as a text file instead of filling the box; see Paste text, screenshots and files. Press Enter to send and Shift+Enter for a new line. When you type with an input method, such as for Chinese, Japanese or Korean, Enter first confirms the word you are composing; it sends only when nothing is left to confirm. The same holds in every Hive box that acts on Enter.
2

Optionally narrow the sources

Type @ to open one menu of agents, skills, Library files and folders, connected tools, Web and built-in public data such as Companies House (see One menu for agents, files and apps). You can also pick a tool or mode, choose an agent from the agent picker, or choose a model. On a phone, the Project and Folder controls under the box show only their icons until you choose one, and a chosen agent’s full name wraps onto its own line.
3

Follow Hive's progress

While Hive works, the line above the answer names the systems it is reading, the time taken so far and, from the moment the answer starts, the model answering, including the one Hive chose when you left the choice to Hive. Hive’s steps are folded under Hive is working: select it to see them, and select it again to fold them away. Until the answer is finished, a line under its text reads Hive is working (or the agent’s name) with the time so far, so you can tell a half-written answer from a finished one. Just above that line, a small chip for each thing Hive is doing right now, such as Reading Gmail, Searching the web or a helper agent, pulses next to a count such as 3 working. The same action on the same source shows as one chip, with ×3 when Hive runs it three times at once, grouped the same way as the steps under Hive is working. Hive’s mark beside the answer pulses gently while it works. The chips go when the answer finishes, or when you stop it. If your device is set to reduce motion, nothing pulses. A screen reader hears one short summary, such as “Hive is working: Reading Gmail, Searching the web”, when what Hive is doing changes. The same chips show in an agent’s chat and in the Studio Test panel. A long answer keeps going if you leave the page or reload, even the moment after you send it. When you come back, the answer appears as far as Hive has written it and keeps growing, then the checked answer replaces it in the same place when it is ready. The steps Hive has taken so far stay in place, and the time so far counts from when you asked, not from when you came back. In workspaces where Hive follows answers live (being turned on gradually), the new text types out smoothly as it arrives rather than a few lines at a time. If your connection drops while Hive is answering, for example when you go offline for a moment or Hive is updated, the answer shows Reconnecting… instead of an error and carries on from where Hive has got to once it can reach Hive again. Stop still works meanwhile. An answer still running after 30 minutes is stopped and says so, so the chat is free again (see Troubleshooting).You can keep working in the chat while Hive answers, including during the first answer of a new chat. The tool picker, the model picker and the agent picker all stay available, and Stop sits beside Ask. Your next question then waits its turn: it shows as Queued, runs after the current answer and starts by itself when the answer ends, with the tool, model or agent you picked for it. See the queued-question notes in Feedback, sharing and history.
4

Read and check the answer

The answer starts with a one-line headline, then numbered findings when there are several. Select N Sources under it to see the exact records it was built from. See Reading an answer.
After a new chat’s first answer, Hive gives the chat a short name in the sidebar, such as “Q4 revenue target”. It writes the name on your workspace’s own efficient model, or the cheapest model your workspace allows when it allows no efficient model, and never on a model your workspace or your access does not allow. If no allowed model can write it, the chat keeps your first question as its name. A name you give the chat yourself is never replaced.

Edit your last question

To correct the latest question you asked, select Edit under it. Hive removes that question and its answer and puts the question back in the box exactly as you sent it: its text, the connected apps you mentioned with @, and the tool you picked, such as Create presentation. The files you uploaded with it are back in the box too, and the chat keeps its agent, so pressing Ask sends the corrected question with all of them. Remove a file or change the tool first if you want to send it without them.

Reading an answer

Every answer has the same parts, top to bottom:
  • The provenance line, such as “Hive · read HubSpot, Google Ads, Xero, Gmail · 6s · GPT-5.5”: the systems Hive actually read for this answer, how long it took, and the model that answered. While Hive works, it says what it is reading and counts the time up, and it names the model as soon as the answer starts. If Hive moves the answer to another model partway through, the line changes to the model that answered. When an agent answered, the line starts with the agent’s name instead of Hive, beside its icon; see Use an agent in Ask.
  • Used skill chips, when a skill shaped the answer: one for each skill you picked, each mandatory skill and each skill Hive chose for the question (hover a chip to see which). A skill you can’t open yourself shows as “a private skill”. A skill you picked that was not used gets a quiet line saying why, such as “Not used: Board pack (Skills is switched off for you)”.
  • The answer: a one-line headline first, then numbered findings when there are several. When a request could mean more than one thing, such as “a short summary of how our Q3 went”, the answer opens with one line saying what Hive took it to mean (which period, which figures), so you can correct it straight away; when a wrong guess would be costly, Hive asks one short question instead. Quick answers say in a few words which meaning they answered. Hive also checks its figures against their sources before it answers. See How Hive handles professional work.
  • Worked (or Worked, thought): folded by default; select it to see each step Hive took — the tool’s icon, what it did, whether it finished and how long it took. Repeated reads of the same system share one row, such as “Read 6 items from Gmail”; select that row to see each read. Steps that need your attention, such as a read that failed, something held back or an action waiting for your approval, always keep a row of their own. When the model gives a summary of its reasoning, it appears as a Thinking step; that is when the label says Worked, thought. The model’s reasoning is never part of the answer itself, whichever model you choose. The steps and the time stay with the answer after a reload.
  • Files, sheets, decks and tables Hive made, as cards you can open, copy or download. A deck Hive made or edited shows a small preview on its card: the first few slides of a new deck, or the slides Hive changed. Select the preview, or Open, to open the deck in the deck editor; in the file panel beside the chat it shows as a live preview of its slides (see A deck in the file panel).
  • Pictures Hive drew, shown in the answer at a size that fits your screen. Select a picture to see the whole image larger; press Esc or the close button to return to the chat.
  • The footer:
    • N Sources shows the icons of the systems the answer cites. Select it to open the Sources panel: one card per system (“Grounded in HubSpot”), listing the exact records, pages and files cited, with links to open them. A system read live says so.
    • Copy copies the answer.
    • Regenerate asks the same question again on the same model, with the same skills you picked for it. The arrow next to it lets you pick another model. It appears under the latest answer.
    • Good response and Bad response rate the answer.
  • A status pill, such as “Approval waiting · needs you”, only when the answer left something for you to do, like an action to approve.

How Hive handles professional work

Hive works to the standard a professional-services firm expects of a careful colleague:
  • Country and date. Hive takes the country from your question, your firm or your files. When none is clear, it says which country it assumed. Tax rates, thresholds and rules come from current official sources, with their date; a blog or news story is never the only support.
  • How sure it is. When the evidence is clear, Hive says so directly. When it is not, it says the point is likely, uncertain or unknown. When sources disagree, it sets out both views and what would settle the question. It works figures out exactly, with a calculation tool when it has one, rather than estimating them, and if it got something wrong earlier in the chat, it says so and corrects it.
  • Advice. Hive’s legal, tax and investment answers support a qualified person’s advice; they are not that advice, and Hive makes no personal suitability recommendations.
  • What it flags. Hive mentions briefly anything worth your attention: a wrong premise in the question, fees or dates that don’t match the engagement letter, non-standard contract terms, a possible conflict of interest, or signs of fraud such as changed bank details or a look-alike sender. When a file, email or web page contains instructions aimed at Hive, it tells you in its own words what they ask, never repeats their links, and never follows them.
  • The other side of a deal. If you ask what your firm did for the other side of a deal or matter you work on (for example, you are on the sale team and ask what your colleagues advised the bidder), Hive answers without searching anything. It flags a possible conflict of interest or information barrier and suggests you check with your compliance or conflicts team. Follow-ups on the same matter, such as “just look it up anyway”, “what did they say?” or a question that names the same company or party again, stay unsearched too, for your next three messages. An unrelated question in between, such as “what’s on my calendar?”, is answered as usual. Asking to read what the other side sent you, such as their emails or offer letter, is ordinary deal work and doesn’t raise the conflict on its own. Once your compliance team has cleared it, say so, for example “Compliance has cleared this”, and Hive searches as usual. If the answer leaves the conflict out, Hive adds its own line saying so at the top.
  • Companies you name. When you ask about a company by name, Hive checks that what it read actually names that company. If the answer cites sources but none of them mentions the company, Hive opens the answer with a warning that the sources are about something else, so another company’s document, your own firm’s included, is never passed off as the one you asked about.
  • Client emails. Drafts put the purpose in the first two lines, keep internal notes and margins out, and never promise a fee, deadline or advice your firm has not approved. If you ask for a reply agreeing to a fee, date or scope that differs from the engagement letter, Hive says so first and drafts it as subject to your firm’s confirmation.
  • Bulk work. Hive starts with one item as a sample before the rest, and suggests extra work rather than doing it unasked.
Hive is open that it is an AI and never claims to be a person. An agent’s own name, and drafts you send in your own name, are not affected. Agents, workflow steps and scheduled tasks follow the same safety rules on fraud, hidden instructions and repeated actions.

Risk tiers

Hive sorts each question you ask into one of four risk tiers before it starts, following its operating standard: The tiers apply to questions you ask in chat, including quick answers. Agents, workflow steps and scheduled tasks keep their own instructions and tools. Hive reads the tier from the words of your question in English; in other languages it follows the same standard from its instructions alone.

The language Hive answers in

Hive answers in the language you write in. Ask a question in Arabic, Spanish or Chinese and the answer comes back in Arabic, Spanish or Chinese; ask in English and it answers in English. When your message does not show a language, such as a file name, a number or a link on its own, Hive uses your Display language from Language and region.
  • Response language in Settings → Account → Preferences overrides both. Choose a language there and Hive always answers in it, whatever you write in, unless you ask for a piece of text in another language. Match the conversation, the default, answers in the language you write in.
  • Names stay as written. People, companies, statutes, legal terms and quotes from your files keep their original language inside an answer.
  • Text for someone else follows its reader. An email reply, a message or a client document Hive drafts uses the language of that thread or recipient unless you ask otherwise.
  • Agents answer the same way in a chat with you. An agent that runs on a schedule, or from a workflow or an app event, has no message to follow, so it writes in the Response language of the member it runs for, or their Display language.
  • Signals are written for you in the same way: card titles, summaries, the reason a card was raised and drafted replies use your Response language, or your Display language. A drafted reply to someone else still follows the language of their thread. A meeting recap sent to several attendees is written once, in your workspace’s default language.
  • Workflow AI steps write in the language their step instruction is written in, unless the instruction asks for another language.
  • Chat titles in the sidebar use the language of your question.
  • “Draft it” works in your language. When you ask for a draft, or say not to send (“no lo envíes”, “لا ترسلها”, “不要发送”), Hive saves a draft and sends nothing, as it does in English.
  • The same safety checks in every language. In English, Arabic, Chinese (Simplified), Dutch, French, German, Hebrew, Italian, Japanese, Portuguese (Brazil) and Spanish, when an answer says an approval card is ready but none was made, Hive replaces that claim with a plain sentence, in the language of the claim, saying no approval card was prepared in this reply. Asking to change a HubSpot record in your language (“Erstelle einen Deal”, “在 HubSpot 中创建交易”) is treated as a change, and a request to see records stays read-only. Naming your mailbox in your own script (“أوتلوك”, “谷歌邮箱”) keeps a draft or email in that mailbox, and asking to rewrite a whole open file (“schreib das ganze Dokument neu”, “最初から作り直して”) lets Hive rewrite it, while any other request changes only the part you name. An email you ask Hive to send, in any language, is still an approval card you decide.
  • Web pages Hive reads for you are requested in your language first, with English after it, so a site with a page in your language returns that one.
  • A few fixed lines Hive writes itself, such as the summary after you approve actions, the note that a card has nothing waiting for you and the label on a sample chart, use your Response language, or your Display language when that is left on Match the conversation.
Answers in languages such as Arabic, Hindi, Thai, Chinese and Japanese use more tokens for the same text than English, so they use somewhat more credits.

Right-to-left languages

Ask lays out Arabic, Hebrew, Persian and Urdu text right-to-left wherever it appears, whatever language the rest of Hive is in:
  • Answers set the direction of each paragraph, heading, list item, quote and table cell from its own words, so an Arabic answer that opens with an English company name still reads right-to-left, and an English paragraph next to it stays left-to-right. This holds while the answer is still being written. Code is always shown left-to-right.
  • Your question follows the language you type in, in the question box, in the sent question and in comments. Every other text box in Hive does the same.
  • Names of files, people, connected apps and cited records keep their own direction, so a right-to-left file name does not move its extension or reorder the words around it. The same goes for names and titles elsewhere in Hive: chats and projects in the sidebar, Library files and folders, Inbox items and notifications, agents, workflows, clients and Signals cards.
  • Phone notifications in Arabic or Hebrew read right-to-left too.
Hive draws Arabic, Hebrew, Hindi (Devanagari) and Thai in matching Noto fonts on every device, with no extra letter spacing pulling Arabic letters apart. If your browser is set to Chinese, Japanese or Korean, Hive loads Noto fonts for those languages as well. Otherwise your device’s own font shows them. Each font downloads only when a page shows its script.

Dates, weeks and time zones

Hive gives dates and times in the time zone you use Hive in. A week runs Monday to Sunday in that time zone, whichever week you ask about: “this week”, “next week” and “last week” follow the same rule in the answer and in what Hive reads. Hive uses your calendar in the time zone it is set to and never asks you to change it first.

When Hive is busy

When Hive is answering a lot of questions at once, a new question can wait for a free slot. The answer then shows Waiting for a free slot… and starts as soon as one frees up, usually within seconds. If none frees up within about 90 seconds, Hive stops and says “Hive is busy answering other questions right now. Ask again in a minute.” A question that waited and never started is not counted against your limits, and Stop works while it waits.

One menu for agents, files and apps

Type @ in the composer to open one menu, grouped by type. Keep typing to search every group at once; skill, agent, app and source names can include spaces, such as @Customer health or @Companies House. Use the arrow keys to move through the list, and press Enter to pick. Escape closes the menu. A pin narrows where Hive searches, and nothing else. With @Outlook pinned, Hive searches your Outlook rather than your other apps, the web, the Business Brain or Signals (pin @Web as well to include the web). Everything else still works on a pinned question: the chat’s files and its project, the Library file or folder you chose, charts, files, pictures, your matters, saving facts and memories, and actions. Pin a mailbox (@Outlook or @Gmail) and Hive can read your email threads in it and reply in the same thread there, never in a mailbox you did not pin. Your own MCP servers are the one exception: pinning an app that isn’t an MCP server keeps their read tools, and pinning an MCP server keeps only the pinned servers’ tools. You don’t have to pin an app to have Hive read it. Name it in your question, such as “chart our FreeAgent profit and loss by month”, and Hive has that app’s own reads to hand from the first step, however many apps you have connected. A skill you pick with @ is followed. Your question keeps the skill’s chip after you send it, the answer shows Used skill: with the skill’s name, and the answer’s text opens by saying which skill it used (a skill of your own is called “your skill” there, never by name). Use in Ask on the Skills page opens a new chat with that skill already picked. Agents use their own skills, so while an agent is selected the skill chips above the box are greyed out with Agents use their own skills and are not sent with your question. Choose Use default Hive to use them again. See How Hive uses skills for the other cases where a picked skill is not used. Hive also uses your active skills on its own: it sees what each one is for and loads one when your question fits it, and the step under Worked reads Using skill: with the skill’s name, or Using one of your skills for one of your own. Mandatory skills always apply. Ask “make me a skill that…” and Hive writes one and saves it to your skills, switched off until you review it. The menu only lists what you can use: agents you can see, and Library files and folders you own or that are shared with you. Library results load when you open the menu and show the closest matches, so type part of a name to find older files. In a chat you started, you can add more Library files and folders after the first message too: they join the chat’s Library context with your next question, and the files already chosen stay. In someone else’s shared chat the menu doesn’t offer them. Mentioning a person works in comments on a shared chat, not in questions to Hive.

Speak your question

Select the microphone in the composer and speak. While you record, the microphone becomes a recording bar: a red dot, a level meter that moves with your voice, and the time so far. Your question stays in view, and your words are transcribed in pieces of at least 10 seconds, so the text appears in it while you talk, a few sentences at a time; whatever is left is transcribed when you stop. Each piece counts as at least 10 seconds of dictation. Select the tick (Insert text) to stop and keep what you said, or the cross (Discard recording) to throw the recording away and put the box back as it was. Recording stops by itself after 10 minutes. If your device is set to reduce motion, the meter is a simple level bar. On Business, dictation uses your own AI credits, at what the transcription costs Hive; if you have none left to use, the microphone stops and says so. On an earlier plan, dictation is included in the plan: each Pro or Max seat adds 5 hours a month to the workspace, shared by everyone in it, and after that dictation uses 1 credit a minute. While you record, the composer shows how much of this month’s included dictation is left. If the included hours are used up and no credits can be used for more, the microphone stops and says so; an Owner or Admin can add credits in Settings → Billing. Good questions are specific and name the business thing you care about:
  • “Which clients have invoices more than 30 days overdue, and how much in total?”
  • “What did we promise Acme in the last three emails?”
  • “Which deals in the default pipeline haven’t moved in two weeks?”
  • “Summarise this contract and list the renewal terms.” (with the file attached)

Your first questions

Right after you connect Microsoft 365, Google Workspace or another mailbox or file store, a new chat shows Getting your data ready, with one line per connection, such as “Reading your files… 312 so far”, “Reading your email…”, “Preparing 40 items for search…” or “Connecting…”. There is no total, because providers do not report one. The card updates every few seconds. When at least one connection is searchable, the card becomes Your data is ready: “Hive has read N files and M emails from Microsoft 365”, with up to three questions built from your own data:
  • one about your most frequent recent correspondent (last 30 days), such as “What did we agree with Acme last month?”;
  • one asking for a summary of your most recently changed document;
  • one listing emails still waiting on a reply.
They are tailored to your firm type (law, accounting, advisory or general) from the website research. A question with no data behind it is replaced by a general one, such as “What does Hive know about our firm so far?”.
  • Selecting a question sends it straight away, scoped to the connector it came from (shown as an @ chip).
  • If a connection cannot be read, the card says so, for example “Couldn’t read Outlook”, with an Open Integrations link to reconnect it. When an account is still connected and Hive has read most of it but a few items keep failing, it says how much it read instead, for example “922 files read; some couldn’t be read”.
  • Hide the card, or use one of its questions, once your data is ready, and it stays hidden for you in that workspace on that browser. While Hive is still reading, the card says you can ask questions already, and its button is Hide for now: the card stays out of the way in that browser tab and comes back as Your data is ready when the read finishes.
  • The card links the quickstart at its foot.
  • On a phone, the card shows its message and Hide but not its questions, so the new chat keeps to the composer and the tools, such as Create document.
The card and its questions are built only from what you could open in Ask yourself: email counts and correspondents come from your own mailbox, and an admin gets no wider view. Only you see them.

Suggested questions

Under the composer, a new chat suggests three questions. The first two are for the tools you have connected: at most one per kind of tool that is live or syncing, such as “What emails need a reply from me?” when Gmail or Outlook is connected (it covers all your connected mail) or “@OneDrive What changed in our docs recently?”. General questions fill any free place, and the last is always an @Web question: “What are people saying about us online?”. Selecting one sends it straight away, scoped to the tool shown. With nothing connected yet, the first two are questions Hive can answer without any of your data: “What can Hive do for a firm like mine?” and “Which tools should I connect first?”, then the @Web question. Once a tool is connected, the suggestions above take over. Connect your tools, along the bottom of the box on a new chat, opens Integrations. A connector with an error gets no suggestion, and while the Your data is ready card shows, the suggestions are hidden. On a phone, the suggested questions and their pencil are not shown, so the new chat keeps to the composer and the tools, such as Create document; edit your starters from a computer or tablet.

Choose your own starters

Make the suggestions your own shortcuts. Select the pencil at the end of the suggestions (Edit your starters) to open Your starters, which starts from the questions you see now:
  • Add a starter gives it a Label, the text on the button (up to 40 characters), and a Prompt, the question Hive is sent when you select it (up to 2,000 characters).
  • Edit any label or prompt, move a starter up or down to reorder, or remove it. You can keep up to six. Every starter needs both a label and a prompt before you can save.
  • Save replaces Hive’s suggestions with yours. Your starters are sent as plain questions, with no connector pinned, and Hive answers them like any question you type. If you remove every starter and save, the suggestions stay empty; the pencil reads Add starters.
  • Reset to defaults brings back Hive’s suggestions.
Your starters are yours alone: they are saved for you in this workspace, nobody else in the workspace sees them, and another workspace you belong to keeps its own.

Where answers come from

Library files and folders you choose for a chat are read whole, not searched in pieces: PDFs and images as the files themselves when the chat’s model can read them, Word files as their full text, PowerPoint decks as their text per slide with speaker notes (and slide pictures for models that read images), and CSV, TSV and Excel files as tables Hive can calculate and chart from. LibreOffice and OpenDocument files are read like the Microsoft files they match. When they add up to more than about 50,000 tokens, Hive lists them and opens the ones it needs one whole file at a time. Every model can work with your photos: a model that can’t see images gets Hive’s description of each image, and of each slide picture, instead. See How Hive reads the files you choose. When Hive lists messages from Gmail, it also reads the sender, recipients, subject, date and a short preview of up to 12 of them, so an answer or an agent step can say who wrote and what about. It reads those headers and the preview only, not the full message body. If one message cannot be read, the answer says so rather than skipping it.

Direct reads (pilot)

Ask reads these connected tools directly instead of first working out which request to make, so their answers come back faster:
  • Gmail and Outlook: search all your mail by words, sender, recipient, who was copied or anyone on the message (so one search finds your mail with a person both ways), dates and unread, newest first, and open one message to read its body and attachment names. Up to 20 messages per mailbox. A question about your email that names no mailbox (“read my sent emails”) searches every Gmail and Outlook account you connected at once, with each message and step labelled by its account, and says which mailbox could not be read; naming one (“in my Gmail”) narrows it. See Ask about all your mailboxes. Hive searches one mailbox at most three times for a question, then works from what it found. Mail you have thrown away stays out of the answer: opening a Gmail thread leaves out messages in Trash or Spam, and opening an Outlook conversation, or searching all your Outlook folders, leaves out Deleted Items and Junk Email. Hive says how many it left out. To read that mail, ask for it (“include the ones in my trash”, “search my junk folder”).
  • Your email threads: list the threads where someone is waiting on your reply (never one you were only copied on, on Cc or Bcc or through a group or forwarding address, and never a mailing or an automatic notice such as an account confirmation, a waitlist or terms update, or a failed-payment notice). One list shows up to 20 threads: when more are waiting it keeps the most recent and the longest waits, and says how many more there are, so you can ask for a date range to see the rest. Ask can also read one whole thread from your own mailbox: every message oldest first, with sender, recipients, date and body, your own messages marked as yours. In a long thread the latest messages are read in full and older ones are shortened. Hive finds threads among the most recent mail it keeps for each mailbox (the newest 500 messages), and says so when earlier messages of a thread may be missing. A thread from a mailbox you did not connect yourself cannot be read, including a colleague’s.
  • Google Calendar and Outlook: list your events between two dates, up to 92 days and 50 events at a time, or only the events one person is in (as an attendee or the organiser), by their name or address.
  • Google Drive, OneDrive and SharePoint: find files by words in their name or text, and read the text of one file (documents, spreadsheets, PDFs). Up to 25 files per search.
  • Microsoft Word: find Word documents in your OneDrive and read the text of one.
  • HubSpot: search deals by words, stage and close date, and find contacts and companies. Each deal, contact and company links to its page in HubSpot once Hive knows your HubSpot account (it learns this when you connect).
  • Salesforce: search across accounts, contacts, opportunities and leads by words, list opportunities by stage and close date, and open one record. Each links to its Salesforce page.
  • Stripe: list payments, invoices and subscriptions, find customers by name or email, and read your balance. Amounts are shown in each currency, and each record links to the Stripe Dashboard.
  • Notion: find pages and databases by their title and read the text of one page (its top-level blocks; nested blocks and database rows are not read).
  • Microsoft Teams: list your teams, their channels and your chats, and read the latest messages of one chat or channel (not replies).
  • Xero, QuickBooks and FreeAgent: list invoices by status and date, up to 50 at a time. For Xero, read the profit and loss report for a period; for FreeAgent, the trial balance summary (the total of each income and cost category), which answers profit and loss questions. Unless you name a period, FreeAgent is read for this month to date and the answer names that period.
Each read uses your own connected account and only reads. It returns a bounded number of records and amount of text. Each read is cited as one source: a list of messages, events, deals or invoices is one source named after the tool and what was asked (for example HubSpot · "Acme" · deals), and a single message or file is cited as itself, with a link back to it where the tool provides one. If an account is disconnected, refused or slow, the answer says what could not be checked rather than saying there is nothing. When you have connected several accounts of a tool, Hive reads up to three; name the one you mean (“my work Gmail”) and it reads only that one. Other reads from your connected tools work as before. Workflow agent steps use the same direct reads for their accounts; see Agent steps. The newer Ask also reads Companies House directly, with no connection: search for UK companies or officers, and read a company’s profile, officers, filing history and persons with significant control, each cited to its page on the public register (one source per page, so a company’s directors are one officers source). For UK company questions it prefers the register over the web. See Companies House. It reads the other built-in public registers the same way: LEIs and parent companies from GLEIF, EU VAT checks from EU VIES, US filings and financials from SEC EDGAR, HMRC guidance and manuals from GOV.UK, French companies from Recherche d’entreprises, Gazette notices from The Gazette, sale prices from Land Registry Price Paid, open tenders from UK public tenders, official statistics from ONS, Norwegian and Finnish companies from the Norwegian and Finnish business registers, and, where your deployment has them set up, the FCA Register, the Charity Commission, the Scottish Charity Register, the Irish company register and Insee Sirene. Each is cited to the register’s own page with the credit its licence asks for, and a workspace admin can switch each one off. Hive only reads what the connected account is allowed to see, and connector data is treated as data, never as instructions. An answer that read anything through your connected accounts is not used to teach the shared Business Brain; see Learning from Ask answers.

Ask about a person

Ask “What’s happening with Stacey?”, “Any news from Tom?” or “Draft a reply to Priya” and Hive gathers that person’s context before it answers:
1

Find who you mean

Hive turns the name you used into the person and every email address their mail is under. It reads your own Business Brain (names, alternative names and email facts — including the people Hive learned from your own mail; see People and companies from your mail), the last 20 messages in your mail that mention the name, and your calendar. An address the Brain holds and the addresses their mail actually uses are joined, and so are short and long forms of one name (“Zach Ong” and “Zachary Ong”). It also lists the threads (each once, however long, newest first) and the meetings it found with them, so Hive can open every one of those threads straight away. Under Worked this shows as Finding who you mean. When more than one person has that name — two Staceys — Hive lists them and asks which one you mean.
2

Read your mail with them, both ways

It searches each mailbox you connected for messages from, to or copied to that person, by their name, which matches every address they use, and reads every thread it found in full, side by side. A search by a person’s address that finds nothing searches again by their name.
3

Check your calendar

It lists your recent and coming meetings with them.
4

Check the Business Brain

Last, it reads what the Brain knows about them: the person, their organisation and the open items and matters linked to them — never your own matters, your company’s alerts or other entities the question happened to touch.
The answer covers only what is linked to that person — where things stand, what is waiting on whom, and the dates coming up — and leaves out anything not tied to them. When Hive finds nothing, it says what it checked. Before it drafts a reply to someone, it reads your earlier mail with them so the reply follows what was said and matches its tone. Hive does this from the first step whenever your question names someone after words such as “with”, “from” or “reply to” — written with a capital letter or as an email address, and in lower case too after “happening with”, “news from”, “hear back from” or “reply to” — and gives such a question a little more room (10 steps rather than 8). For other questions about a person it still follows the same order. It reads only your own mail, calendar and Brain — never a colleague’s. Finding who you mean is not offered under an @ pin (a pin narrows Ask to the sources you named) or to agents. Each person question adds a few steps to the answer: finding the person, then the mail and calendar reads. Finding the person counts as one of the three searches Hive makes in a mailbox for a question.

How Ask picks its tools

Ask doesn’t hand the model every tool you have on every step. When a question has more than 20 tools to choose from — Hive’s own, your connected apps’ reads, Hive app data and your MCP servers’ tools — the model starts with a small set of everyday tools: searching your documents and attached files, the web, the Business Brain and your connected apps, making or editing a file, a chart, and proposing an action for your approval. Up to 20 of your own MCP servers’ read tools are in that starting set too, so Ask uses a server you connected, such as Carta, without searching for it first. A question about a person also starts with your mail and calendar reads and Finding who you mean (see Ask about a person). It’s also told which apps and servers it can find more tools for, and searches for the others when the question needs them. A tool it finds is ready from its next step, and stays ready for the rest of the answer. Under Worked, that search shows as a Finding tools step. The search only ever looks at tools you can already use in that question: what your plan, your workspace’s settings, your own connections, your @ pins and an agent’s tool list allow. It never finds a tool you couldn’t use otherwise, and a tool it finds asks for approval exactly as before. One step offers the model at most 60 tools.

Reading whole web pages (pilot)

Web search returns short snippets, which often stop before the detail you need: a full list of a company’s directors, a table or a filing date. On the newer Ask, Hive can open the page itself and read it in full. For example, ask “Who are the directors of HIVE AI LIMITED, company number 17478717 at Companies House?” and Hive searches, opens the company’s Companies House page, follows its People link and lists every officer, citing that page. Hive only opens:
  • a page a web search returned while answering this question;
  • a link you wrote in your own messages in the chat; or
  • a link on a page it has already opened, on the same site (up to 40 per page).
It cannot open any other address. Reading your workspace’s data first does not change this: web searches and page reads are never held back by Hive’s privacy rule. Each question can open up to 6 pages. Hive reads public HTML pages, plain text and PDFs over HTTPS only, never private or internal network addresses. Each page is limited in time and size, and it reads up to 24,000 characters of English text from one page (fewer characters of a language that takes more tokens per character, such as Chinese or Arabic). If a page is longer, the answer says only part of it was read. If a page cannot be opened, for example because the site refuses automated readers, the answer says so and uses the search snippet instead. Page text is treated as data, never as instructions. Web pages are cited like any other web source.

Citations

Every answer can cite up to 20 sources. The model only points at evidence; Hive fills in each citation’s label and link from the real record, so a citation cannot be invented. Citation links open over https only. Citations can point to web pages, documents, emails, live tool reads, signals, Brain facts, memory, approvals and sections of these docs. Each source names where it came from, the same way on the web and in Slack or Teams. In the Hive app, the Sources panel groups them into one card per system and lists each cited record under it: Only a live read is labelled with a connector’s name: a Brain fact or a signal that mentions a customer or a tool is still labelled Brain or Signals. Each page appears once. A list read is one source rather than one per record, and the same page read twice in an answer keeps one source. A search Hive made only to find something, such as a company number, is left out when the answer also cites a page from the same tool. When the model does not point at its evidence, Hive chooses up to five sources itself from what it read for your question:
  • It ranks everything it read, Business Brain facts included, by how many words each item shares with the answer and, less strongly, with your question, and cites the closest matches. A source that matches far less than the best one — often sharing little more than a customer’s name — is left out.
  • When a live read, document or web page matches as closely as a Brain fact, the direct read is listed first, because it is the fresher source.
  • A read that failed or came back empty — a tool that timed out, was disconnected, refused access or returned nothing — is never chosen as a source.
  • Hive then usually rewrites the answer from only the sources it chose.
  • Matching works in languages written without spaces between words, such as Japanese, Chinese and Thai.
If nothing Hive read matches the answer, it shows no Sources rather than attaching ones that do not back it up. Treat an answer without sources as unverified: pin the right tool with @, or attach the document, and ask again.

When a live source cannot be read

If a connected tool could not be reached during your question, Hive tells you at the end of the answer:
Live sources were not checked: … could not be read this turn, so this answer is based on stored Brain data that may be out of date.
Treat that answer as a starting point, and check connection health if it keeps happening. When one of your accounts is disconnected (its access was revoked or expired, or it shows Reconnect required), Ask does not answer as if the data does not exist. If your question needs that source, the answer says so, for example “Outlook is disconnected, so I could not read it — reconnect it on Integrations.”, then answers from whatever else it could read. Ask only reads and acts through accounts that are fully connected: an account still connecting, or one that needs reconnecting, is not used until it is live again. If you @-pin a tool whose only account is disconnected, Ask shows the reconnect prompt instead of reading. The answer can also offer to reconnect it there and then. If you @-pin a tool and Hive cannot work out which records to read from your question, Ask does not guess an answer from other data. It says the tool “is connected, but I couldn’t answer from it for this request” and suggests naming a more specific object or filter, for example “open deals in HubSpot closing next month”. Ask again with the records spelled out. An authorisation failure during an Ask read counts towards connection health, the same as a failed sync, but only once Hive has confirmed the connection itself no longer works: then the account is marked Disconnected and its owner is notified once. A tool that refuses one read is not disconnected for it: until Hive has checked, Ask says it couldn’t read the tool and that the account may be missing a permission or the connection may need reconnecting. When a read of a pinned tool fails, Ask says which tool and why: it couldn’t read the tool (a missing permission or a connection that needs reconnecting, when Hive has not confirmed which), the tool must be reconnected (only once Hive has confirmed its sign-in no longer works), the tool is limiting requests, the tool was unavailable, or Hive could not work out which records to read. The same reasons reach agents, workflow steps and AI clients that use Hive through its MCP server. Some tools are read only through a fixed set of reads their connection allows: LinkedIn, Microsoft Power BI, Tableau, Statista, SurveyMonkey and Vanta. When Ask looks something up in them, it never tries another read, so it does not ask a tool for data its connection cannot return. LinkedIn, for example, reads only your own profile, never posts or connections. Their dedicated read tools, such as reading a Tableau view’s data, keep working. Hive has no verified reads yet for DATEV, so Ask says it cannot read it rather than trying.

Connect or reconnect an app from the answer

This is rolling out to workspaces gradually, so you may not see it yet.
When Hive tried to read an app for your question and could not, because you have not connected it or your own account for it needs reconnecting, the answer can end with Connect an app to answer this: “Hive couldn’t read these apps for this answer. Connect them, then ask again.” The same card appears when you name an app in your message, such as “see what we have in our Calendly account”, and Hive checks your connected apps for it while answering (the Checking your connected systems or Finding tools step under Worked) and finds no account of yours for it: Hive does not have to read the app first. The card shows one button or link for each app to connect or turn on, and one button for each of your accounts to reconnect:
  • Reconnect and the app’s name, when your own account’s sign-in stopped working. When Hive knows which account it is, the button names it too, such as Reconnect Gmail (you@company.com) or Reconnect HubSpot (Portal 1234567), so with several broken accounts of one app you can tell them apart. Each button repairs that same account, as Reconnect on Integrations does, so every workflow and agent that uses it keeps working without choosing it again.
  • Connect and the app’s name, such as Connect HubSpot, when you have not connected the app. The provider’s sign-in window opens there and then, as it does on Integrations.
  • Connect and the app’s name in Integrations when the app needs a choice first, such as a region or edition, or when Hive could not load its list of apps. It opens Integrations searched for that app.
  • Turn on and the app’s name in Integrations, such as Turn on Calendly in Integrations, when your workspace allows the app but an Owner or Admin has not turned it on (enabled it) yet, and you are an Owner or Admin. It opens Integrations searched for that app. Nothing is turned on until you connect the app there, which turns it on for the workspace.
  • Anyone else sees no button for that app, only a line such as “An Owner or Admin needs to turn on Calendly before you can connect it.” Ask an Owner or Admin to turn it on.
The line under the title says what you can do with the card. When every app on it is one to turn on, it reads “Hive couldn’t read these apps for this answer. Turn them on by connecting them in Integrations, then ask again.” When an Owner or Admin has to turn on every app on it for you, it reads only “Hive couldn’t read these apps for this answer.” Hive does not finish the answer for you afterwards: once the app is connected, ask again.
  • Only when Hive tried or checked. The card appears only when Hive tried to read the app during this answer and found it not connected or needing a reconnect, or checked your connected apps for an app you named in this message and found no account of yours for it. A check counts only for an app named in the message you sent, not one named earlier in the chat or only by Hive, so an answer that only mentions an app never carries the card. Neither does a read the app refused for a missing permission: fix that permission in the app.
  • Up to three. Reconnects come first, then apps to connect or turn on.
  • Only you see it. People you share the chat with see the answer without the card, and an agent’s answer in Ask never carries it. It is also hidden if your role cannot open Integrations.
  • Only apps your workspace allows. Ask offers an app only when your workspace allows it, and never WhatsApp Business, whose terms Hive shows when you connect it on Integrations. Connect and Reconnect are offered only for an app an Owner or Admin has turned on, by connecting it on Integrations. When you have not connected an app that is not turned on, Owners and Admins get Turn on, and anyone else a line saying who can turn it on; an app that is not turned on is never offered Reconnect. Ask never turns an app on for the workspace, not even for an Owner or Admin. See Enabling a provider for members.
  • Everyday names. For a few apps whose names are also everyday words, such as Monday, Zoom or Box, Ask offers Reconnect for your own broken account but never Connect or Turn on.
  • Buttons that are no longer needed disappear. Once the account is working again, or you already have a working account for the app, the app is gone from the card. If none is left, neither is the card.
  • Your mailbox. If your only Gmail or Outlook mailbox needs reconnecting, the answer offers Reconnect Gmail or Reconnect Outlook with the mailbox’s address, such as Reconnect Gmail (you@company.com), instead of Connect your email.

When a step fails or is refused

When one of Ask’s steps fails or is turned down, for example a draft that needs a detail it got wrong, Ask fixes what it can from your own words and tries once more, or tries another way, before it answers. If it still cannot do it, the answer says plainly what it could not do, without asking whether to try again. Ask only offers what it can do in your chat. To be reminded or to have Hive check back at a set time, ask for a one-time task (“remind me on 14 Oct at 14:45”); when a chat implies a follow-up, Hive may suggest one as a card you confirm, and never sets one up on its own. For something that should happen again and again, ask for a scheduled task. To be told when something happens, ask Hive to keep an eye on it.

When Hive finds nothing

When Ask cannot find anything relevant, the answer lists the sources it checked, for example “I checked Gmail, your connected files and the Business Brain and found nothing relevant.” It also names any source it could not read and why: it was disconnected (reconnect it), it may be missing a permission or need reconnecting, the tool is limiting requests or was unavailable, did not respond in time, was only partly read, cannot be searched for that kind of question, or was not read this turn. Up to six sources are named, then “and N more”. When a source was an app you have not connected, or your own account for it needs reconnecting, the answer can also offer to connect or reconnect it. A source the question’s scope ruled out — the Business Brain in a Library-scoped chat, for example — was never searched, so it is not named. When an answer rests on stored data while a live read of a connected tool found nothing, it ends with a line such as “I read Gmail live and found nothing matching; this answer comes from stored Brain data.” The line names only the stored sources the answer cites; an answer that rests on the empty live read itself (“nothing from Acme this week”) gets no such line. If the live read itself failed, you see the Live sources were not checked warning instead, never both.

Ask about your Hive workspace

Ask can look up how your Hive workspace itself runs, with your own access, and answer from it: “Did last night’s invoice workflow succeed?”, “Which agent runs failed this week?”, “What is scheduled for tomorrow?” or “Which groups am I in?”. It reads the same information the Hive app shows you:
  • the workflows you can see, their settled run history, and runs still in progress;
  • workflow and agent schedules, and your scheduled tasks;
  • the Agent Studio agents you can see, and the status of their recent runs (status, source, attempts and times — not run output);
  • workspace groups (the public ones and yours; Owners and Admins see all of them);
  • which Hive products you can use, the workspace’s general settings and chat sharing setting, and, for Owners and Admins, its model configuration;
  • your own role and your own notifications.
These lookups are read-only and use your current role: Ask sees nothing here that you could not open in the app yourself. Each one is cited with a label that starts Hive · and names what it read. A lookup that fails or is refused is never cited; the answer says what could not be read instead. Limits:
  • Up to four lookups per question, and up to 30 a minute for each member.
  • Ask cannot look up the member list, billing, the audit log, connected accounts, or other chats. Signals, the Business Brain, approvals and the Library are answered by Ask’s own sources as before.
  • After Ask has looked up workspace data in a question, that answer can still search the web: a web search may use what the lookup found, and the search provider receives that query. A web page is opened only from search results, your own links and same-site links on a page it opened. See What can leave Hive after a private read. A picture you ask for is drawn by Hive’s image model, which may use what Ask found, as the chat model does. You can still ask Hive to act: it plans the action from what Ask already found, without reading your connected tools first, and the account’s level decides what happens, as for any action you ask for.
  • Lookups are offered in a regular chat you ask in the Hive app, including in Web search mode or with @Web pinned, when you can use company knowledge, and to an agent that declares Hive data under The person’s own data, with your access. They are not used when you @-pin a connected tool or a Library folder, in a shared chat, from Slack or Teams, in scheduled summaries or workflow agent steps, or by an agent that does not declare Hive data.
Owners and Admins can turn lookups off for the whole workspace with Hive data lookups in Settings → Chat capabilities. You can turn them off for yourself with Hive data lookups in Settings → Preferences. Both are on by default.

Ask how Hive works

Ask also answers questions about Hive itself from these docs: “How do I connect an MCP server with my own token?”, “What does Draft mean for an account?”, “What happens when our credits run out?”. It searches the published docs, answers from the sections it finds, and cites each one with a link to its page and heading on docs.get-hive.ai. When the docs do not cover a question, the answer says so.
  • Where to start: choose Ask about Hive in the workspace menu, which opens Ask with “How do I …” ready to finish. On any page of these docs, Ask Hive about this page in the page menu opens Ask in Hive with that page’s title already in the question.
  • What it reads: only the public documentation, the same for every workspace. It reads nothing from your workspace, so it works whatever your role or products, in a chat with a pinned tool, and in Slack and Teams direct messages. Agents you build in Agent Studio do not use it; they answer from their own knowledge.
  • Web search afterwards: reading the docs does not stop the answer from searching the web.
  • How current it is: Ask reads a copy of the docs that is refreshed from time to time and ships with Hive’s releases, so a very recent docs change may not reach Ask yet.

Attach files

Select + under the box and choose Add files, or drag files into the composer. Hive reads attached files to answer you. To open a Word, Excel or PowerPoint file as an editable copy straight away, without a question, choose Open a Word, Excel or PowerPoint file to edit (see Tools and modes). An Excel file opened this way keeps its number formats, column widths, wrapped text and alignment in the sheet editor, and a first row of data stays data rather than becoming a row of column letters. Select from Library lists the Library files you can open; type part of a name to search them, and pick one to add it to the chat’s Library context. It is not copied: the chat reads the Library file itself, with your access, and anyone you share the chat with still can’t open it unless the Library file is shared with them. You can pick several, one after another, in a new chat or in one you started. The same menu has Select from Google Drive, Select from OneDrive and Select from SharePoint, each with its own logo, for the ones you have connected (see Files from Google Drive, OneDrive and SharePoint), and one Connect link to Integrations naming the ones you have not, such as Connect OneDrive or SharePoint. Comments on a shared chat can’t take attachments at present: in Comment mode the composer has no +, and you attach files in Chat as usual. A file over 25 MB is read from the text Hive extracts when you attach it, not from the file itself: the answer gets all the text Hive could read, but no page images, pictures or colours, and says so. A very large file may be read only in part, and the answer says what it read: a text file or web page over 16 MB is read from its beginning, a CSV or TSV over 16 MB from its first rows, and a large PDF made mostly of pictures as the text Hive could find in it. Attach a file over 50 MB on its own, one per upload. If many large files are being uploaded at the same moment, Hive may ask you to try again in a moment; nothing is added until it succeeds. Ask reads the types in the table. A file of any other type is kept with the chat as a download: Ask doesn’t read what is inside it, and it is never opened in your browser. Text formats such as JSON, XML, YAML and log files are the exception: when the file is plain UTF-8 text, Ask reads it like a .txt file. Hive refuses a program, script, shortcut or disk image, and a program renamed to look like another file. See Supported file types. Macro-enabled Word, Excel and PowerPoint files (.docm, .xlsm, .pptm, and the templates .dotm, .xltm and .potm) are stored without their macros, as in the Library, and Ask reads the macro-free copy like any other Office file. See Macros are removed.

LibreOffice and OpenDocument files

You can attach files saved by LibreOffice or another OpenDocument app: documents (.odt, .ott, .fodt), spreadsheets (.ods, .ots, .fods), presentations (.odp, .otp, .fodp), drawings (.odg) and formulas (.odf). Ask reads them like their Microsoft equivalents: a document’s headings, paragraphs, lists and tables; a spreadsheet’s sheets and cells, with each formula read as the value the file was saved with; a presentation one slide at a time with its speaker notes; a drawing’s text on each page; and a formula’s text. They preview like Word, Excel and PowerPoint files. They are read-only: Hive does not open them to edit, so Open a Word, Excel or PowerPoint file to edit doesn’t take them. See LibreOffice and OpenDocument files.

Send files without a message

You don’t have to type anything to send files. Once at least one attached file is ready, the send button works with the box empty, in a new chat, an existing chat or a project chat, on a computer or a phone; the same goes for a file you picked from Google Drive, OneDrive or SharePoint. Hive says what each file is, sums up what matters in it, and asks what you would like done with it, in the language you have been writing in (or the files’ language, in a new chat). Your message shows only the files. A new chat is named from Hive’s reply, as usual.
  • While a file is still uploading or being read, the button waits, as it does with a question.
  • With nothing typed and no file ready, the button stays off: Type a question or attach a file to send.
  • Regenerate and Retry re-ask the words of a question, so a message of files alone doesn’t offer them. Ask a follow-up instead, such as “Summarise the second file”.

Older Excel files (.xls)

You can attach an older Excel 97-2003 workbook (.xls), in a chat or to a project’s files. Hive converts it to an Excel Workbook and stores that .xlsx copy under the same name, so Budget.xls appears as Budget.xlsx, and from then on Ask reads it like any other .xlsx: every sheet, its cells and its tables. When you attach one, the composer tells you: Budget.xls is Excel 97-2003 — Hive converts it and stores an .xlsx copy. The .xls itself is not kept.
  • An .xls can be up to 25 MB.
  • The file must really be an Excel 97-2003 workbook. A file only named .xls — for example a CSV or web page some systems export with that extension, or a Word document — is refused with “Report.xls” is not an Excel 97-2003 workbook. Rename a CSV to .csv and attach it again.
  • Macros in the workbook never run, and the .xlsx copy has none.
  • Converting takes a few seconds. If Hive can’t convert the file right now, or it takes too long, Ask says so and stores nothing: try again in a minute, or open the file in Excel, use Save As to save it as an Excel Workbook (.xlsx) and attach that copy.

Paste text, screenshots and files

You can also paste into the box, as in other AI chat apps:
  • Short text goes into the box as plain text, without its formatting.
  • Long text — more than 2,000 characters or more than 40 lines — is attached as a text file named Pasted text 1.txt (then Pasted text 2.txt, and so on) instead. Its chip under the box reads Pasted text 1 · Text · 312 lines: select it to read the text, or select × to remove it before you send. Pressing Ctrl+Z (⌘Z on a Mac) straight after the paste removes it too. Hive reads it like any other attached file, and it is listed with the message’s files once sent.
  • Screenshots, pictures and copied files are attached as if you had chosen them with +: several at once, with the same limits, error messages and upload progress.
  • A mix, such as a picture copied from a web page together with its text, attaches the files and treats the text as above.
Paste files while an earlier upload is still running and the composer asks you to wait for it to finish, then paste again. Where a chat can’t take files — a chat with an agent that reads only the sources its setup names, for example — long text is pasted into the box. Each attached file shows as a chip with a preview: a PNG, JPEG or WEBP picture shows a thumbnail of itself, cropped from its top-left corner so a screenshot’s heading stays visible, and any other file shows its type, such as PDF or DOCX. The chip is the same under the box before you send and on your message once it is sent. Select a chip to open the file and check what it is, the same way it opens from Chat files (see below). Files you attach stay with the conversation. To reuse files across many chats, put them in a project or the Library. In a shared chat, everyone it is shared with can open the files you have sent in it, read-only; see Feedback, sharing and history. In a chat someone shared with you as Can collaborate, you can attach your own files to your questions too. They stay yours: Hive reads them for your questions only, and once you send one, everyone in the chat can open it, read-only. Can view members get no attach button. A question that waits in the queue still can’t carry files. Opening a Word or PowerPoint file from Chat files closes the list and shows the file’s pages inside Hive, as it looks in Office, the same way the Library does (see Pages and text). The pages are made when you attach the file; until they are ready the viewer says Preparing pages…. A file over 25 MB, or one whose pages can’t be made, says Pages could not be made for this file. and offers Download instead. PDFs and pictures still open in a new tab, and other files download. When you remove an attached file, or delete the chat, Hive permanently deletes the stored file within about 20 minutes: the copy in storage, the text Hive read from it and its search index. If the same file is still attached to another chat, or is in a project or the Library, it is kept until the last of those is deleted. There is no undo. See Deleted files.

How Hive reads attached files

Files you attach to a question are read whole, the same way as Library files you choose: A file over 25 MB is read from the text Hive extracted when you attached it, and the answer is told what that covers, for example the text of all 120 pages but no page images or pictures. A file Hive cannot read, such as one its safety checks held back or one with no text over 25 MB, is named in the answer with the reason. Any model in the picker that reads images (a vision model) gets an attached JPEG, PNG, GIF or WebP picture as the image itself, so it can describe what it shows. A model that does not read images is told the file is a picture it cannot read, and that a model which reads images can. Long attached files, a long paste attached as Pasted text 1.txt among them, are read like long Library files, with the same limits: past about 50,000 tokens in total Hive first sees a list of the files, then opens each one it needs, and a file longer than about 100,000 characters of English (about 25,000 tokens, so fewer characters of Chinese, Arabic or Thai) comes in parts. When your message comes with an attached file and asks about all of it — “read it from start to end”, a summary, a list, a count, a comparison, or whether something is in it at all — Hive reads the attached files first and works through the parts in order from part 1, rather than searching the file for a few passages. Worked shows Reading the attached files, then each part it opened. A question about one passage, such as “what does it say about penalties?”, can be answered by searching the file. See How Hive reads the files you choose. To have Hive fill in or change an attached Word, Excel or PowerPoint file, just ask: it opens the file as an editable copy in the chat’s file panel, makes the change there and gives the file back in its own format. See Fill in or change a file you attach.

Totals, averages and counts

When a question needs a number worked out from a spreadsheet — a total, an average, the largest or smallest value, or a count — Hive calculates it over every row of the table rather than adding rows up in the model, so a 5,000-row workbook totals to the cent. Worked shows the step as Calculating from the table, and the answer quotes the numbers Hive calculated.
  • It works on CSV, TSV and Excel files and Google Sheets you attach, pick from Google Drive, OneDrive or SharePoint, or choose from the Library; on spreadsheets Hive finds in your documents; and on the records a connected app read returned, such as your invoices.
  • It can sum, average, find the minimum or maximum of, count, or count the distinct values of one column, for the whole table or grouped by another column: “What is the total Amount by Region?”
  • Numbers keep the column’s own decimals, and an average keeps two more. A grouped answer lists up to 200 groups, largest first, and says how many more there are; the all-rows figure always includes every group.
  • It works whether or not your workspace allows charts. Agent Studio agents and workflow agent steps that read spreadsheets calculate the same way.
  • It can first keep only the rows you ask about: “What is the total Amount by Region where Status is Paid?” A filter keeps rows whose column equals one of the values you name, ignoring case and extra spaces, and several filters must all hold. The answer says which rows it kept and how many. If no row matches, Hive says so and names values that column does hold, rather than answering zero. Filters match exact values only, not ranges such as “over 1,000” or dates between two days.
A chart and its answer agree: when Hive charts a spreadsheet, the answer quotes the values the chart drew.

Files from Google Drive, OneDrive and SharePoint

If you have connected your own Google Drive, OneDrive or SharePoint, you can add one of its files to a chat without downloading it:
  • Type @ and pick it from the Google Drive and OneDrive group, or select + and then Select from Google Drive or Select from OneDrive, which lists just that drive’s files. Before you type, the list shows your most recently changed files and says so; type part of a name to search. When you opened it from + and nothing matches, it says No matching files rather than closing. A long list scrolls, and the menu says when there is more below.
  • Each file shows the day it was last edited, its type and the account it is on, so two files with the same name read apart. Two files with the same name, type and account edited on the same day also show the time.
  • OneDrive’s list includes only files on your own OneDrive, newest change first, including files in folders. A file someone shared with you that lives on their drive is not listed, because Hive reads picked files from your own drive.
  • For SharePoint, select + and then Select from SharePoint. Choose a site, then one of its document libraries, then open folders until you reach the file; Back goes up a level. Type in the search box to find a site by name, or, inside a library, a file anywhere in it. The file is added as a chip like a Drive or OneDrive file. If your SharePoint is not connected, or your workspace does not allow it, the picker says so, with a link to Integrations.
  • If one of your accounts can’t list its files at that moment, the picker names it, so a missing file is explained. Try again, or reconnect it in Integrations.
  • Google Docs are read as Word documents, Google Sheets as one table per sheet, and Google Slides as text per slide with their pictures. Word, Excel, PowerPoint, PDF, CSV and text files, from any of the three, are read the same way as when you attach them.
  • Hive reads the file through your own connection, as you, so it sees only files your account can open. It never reads one through someone else’s connection.
  • If Hive can’t read a picked file (it was moved or deleted, it is too large, or it is a kind of file Hive does not read), the answer says so. The file is not counted or cited as a source.
  • Picked files stay with the chat for follow-up questions until you remove their chip, up to 5 at a time. Each is read whole, up to 25 MB; download a larger file and attach it instead, up to 100 MB.

Create things, not just answers

Ask can produce editable Word documents, PowerPoint decks, Excel workbooks, charts and images, and run web research with citations. Pick the tool from the composer, or, in your own chat or one shared with you as Can collaborate, ask for the file in your own words and Hive makes it, for the kinds your workspace allows — see Tools and modes. A new Word document comes laid out, with a title block, figures in tables and native charts, a header and page numbers — see How Hive designs a new document.

Hive files and files in OneDrive or SharePoint

A Word document, Excel workbook or PowerPoint deck you ask for is a Hive file: it lives in the chat, where you can open, edit and download it. “Create a Word document called Acme minutes” makes a Hive document, and the answer says so (“I made a Hive document …”). It never says a file was created in your OneDrive or SharePoint when it was not. To have the file made in your OneDrive or SharePoint instead, say so in your request: “Create an Excel file called Aged debtors in my OneDrive”, “Save the proposal to the Clients folder on OneDrive”, “Make the engagement letter in the Finance SharePoint site”. If your Microsoft account is connected, Hive prepares an approval card to create, copy, upload or share the file there, and the answer says it is a card (“Here’s a OneDrive card to create …”). The file exists in OneDrive or SharePoint only after you approve the card. If no Microsoft account is connected, Hive makes a Hive file and tells you that connecting OneDrive or SharePoint lets it make the file there. In the deck editor, Hive AI can take a deck’s look from a PDF you attach, such as a brand book or lookbook, however large the file (up to the upload limit). It reads the pages that say most about the design (the cover, the first and last pages, pages with a logo in a corner, the most colourful pages and pages that name colours or fonts) and tells you which pages it looked at, out of how many, with the colours as hex codes, the fonts and a few of the file’s pictures. It can then rebrand the deck in the file’s own colours. The first time, a large file is read in the background, so Hive may say it is still reading it; a very long file is read up to its first 500 pages, and Hive says what it could not read, such as the colours of an encrypted file. A new presentation can start in such a PDF’s design too. Attach the file (or keep it in the chat’s project files) and ask for the deck in its brand or look, for example “Make a pitch deck in the style of the attached brand book”. Hive reads the file’s design the same way before it writes the deck, so the new deck starts in the file’s colours and fonts, and the answer tells you which pages it looked at, out of how many, and what it could not read. A font that not every Office install has is replaced by the closest one that does, such as Arial for Helvetica, and Hive says so. If Hive is still reading the file, it makes the deck in one of its own themes and says so; ask again shortly to restyle it in the file’s design. When a slide needs a picture, Hive AI in the deck editor or the deck’s chat can draw one with your image model, when image generation is on for you. It is charged like any other AI image, stored in the deck and placed on the slide, and Hive AI tells you the picture is AI-made. Hive can also put your own pictures into a deck, in an existing deck or a new one (“use my photo as the cover”): one you attach or paste, one you sent earlier in the chat, or one from your Library, the chat’s project, Drive or OneDrive. It uses only pictures you can open yourself and does not charge them as AI images.

Ask Hive to do something

You can ask Hive to act: “Draft a reminder to Acme about invoice 1042”, “Move the Globex deal to Negotiation”. Depending on the level of the account it would use, Hive takes the action and tells you what it did, or prepares it as a proposal and shows it inline in the conversation.
  • It passes Hive’s policy gate: the kill switch, the autonomy ladder, spend caps and role limits all apply. See Autonomy and blast radius.
  • When the action comes to you as a card, review its details before anything runs, then choose Approve or Decline. If an approved action fails and can be retried, the card offers Retry.
  • Nothing on a card happens until you approve it. A draft you asked for is saved in Gmail or Outlook only when you approve its card, so Ask says that plainly. If an answer claims a waiting action is already done, such as “I’ve updated the draft in your Gmail thread”, Hive replaces that claim with a plain sentence: “Nothing has been saved in Gmail yet: approve the card to save the draft there.” Hive recognises these claims in English.
  • The answer describes only what its cards do. If it says a card does more than the card shows, such as inserting a column when the card only writes values into cells, applying a label the card only creates, attaching a file the card has no attachment row for, or sharing a file with no sharing card in the reply, Hive takes that claim out and ends the answer with what each card does, for example “Prepared in this reply: Gmail: create a label (Label: Q4)”. When the email text on a card says a file has been shared or is attached and no card does that, the answer says so; the text itself stays as written, and you can edit it before you approve. Hive recognises these claims in English.
  • When you ask Hive to insert a column or row in a spreadsheet, it never prepares a card that writes over the cells where the new column or row should go. Writing values replaces what is in those cells, so Hive prepares an insert or says it cannot insert there. When the insert comes first, Hive can follow it with a card that fills the new cells, such as the new column’s header. Adding a row below the last row of a table is not an insert, so Hive writes into the empty row.
  • The card shows the full text it will send or save, such as an email reply, with its line breaks. A long text starts collapsed: select Show full reply (or Show full text) to read all of it. Hive does not propose a text longer than 20,000 characters, so the text on the card is always the text that is sent.
  • On a card that sends or saves an email from Gmail or Outlook (a reply, a draft reply, a new message or a draft), you can change the text before you approve it. Select Edit reply (or Edit draft reply, or Edit message) on the text row, change the words, then choose Approve. Select Cancel edit to put Hive’s text back. Only the text changes: the recipients, the account, the thread and the subject stay as the card shows them. The counter shows the limit: 4,000 characters for an Outlook reply, 10,000 for a Gmail reply and 20,000 for a new email or draft. Hive sends exactly your text, and once it has gone out the card shows it as yours. If the card changed after you checked it, Hive sends nothing rather than the original. Declining never sends your edit. Editing is on cards with one action. On a card that groups several actions, decline the row and ask Hive again.
  • The connected account must allow it: without an agent, Hive sends and changes nothing through an account on Read. See Accounts and autonomy.
  • The account’s level decides: through an account on Read Hive prepares nothing and says the account is set to Read; on Draft you get the card; on Act Hive takes the action and the card shows it as done, with a pointer to undo it from the audit log where it can be undone. An action that cannot be undone, or one the safety checks hold back, still comes to you as a card. Bulk actions, saving a file to Drive and tools on MCP servers always ask first, and so does every email Hive sends for you: it comes as a High risk card on Act as on Draft. Saving an email as a draft is not affected. See Read, Draft or Act.
  • One request can change more than one app, such as “email Sam the agenda from Outlook and put the meeting in my Google Calendar” or “share the deck from Drive and draft an email about it”. Hive prepares a card in each app, for up to three apps in one reply, and never the same card twice. When Hive has to read something first, such as the document you named, it reads it and then prepares the card in the same reply instead of asking you. The answer names only the cards that exist: if a card for one of the apps could not be prepared, Hive says so plainly rather than describing it as done or waiting. An email whose text still holds a placeholder, such as “[insert the sharing link]”, is never prepared: a sharing link exists only once you approve the share card, so Hive does not write it into an email in the same reply.
  • When the action creates something new (an email draft, an event, a contact) and you have more than one account for the app, the card shows a Send from picker and you pick the account before you approve. Name the account in your message (“from my Outlook account jo@work.example, draft an email to …”) and the picker starts on it; pin one account with @ to choose it outright. See Which account something new is created in.
  • If you have more than one account for the app, for example two Gmail mailboxes, the levels decide which account Hive uses. With exactly one on Act, Hive uses it and the card names it. With more than one on Act, or several on Draft and none on Act, the card shows a Send from picker. Pick the account, then Approve. On a card that groups several actions, expand a row to change its account. See Which account an action uses.
  • For an app whose actions only ever create something new (a LinkedIn post, for example), Hive only checks that the account is connected and yours before it prepares the card. For other apps Hive may read first, for example to find the deal you want to update. A read the app refuses no longer cancels the action: Hive still prepares whatever needs no read, such as a new calendar event, CRM contact, document or email draft, and tells you which app it couldn’t read and why, whether or not it prepared anything. For an app it couldn’t read, Hive prepares a card that creates something new only when your own message asked it to create something (“create”, “add a …”, “a new …”, “draft”, “schedule” and the like) and asked for no change, so a change you asked for is never swapped for a new record: otherwise it tells you it couldn’t read the app instead. In a workflow or agent run no such card is prepared for an app Hive couldn’t read. A card that changes, removes or sends something is not affected. A read refused for one record never counts against an account that returned other data for the same request. If Hive confirms during the request that an account’s connection no longer works, it prepares no action that would go through that account, including an email sent from that mailbox; cards for your other accounts of the same app stay.
  • On the newer engine, a reply to one of your email threads goes in that same thread, from the mailbox it arrived in, to the latest person on the other side (at their Reply-To address if they set one): Outlook’s own reply, or a Gmail reply filed in the thread with the headers that keep it threaded for the recipient. When you name a particular message (“reply to Nina’s message of 1 November”), Hive answers that message rather than the latest one. Ask to reply all and the reply also goes to everyone else on that message: its other To recipients stay in To and its Cc in Cc, and the card lists each of them under Also to and Cc before you approve. Hive takes the recipients and subject from the stored message, not from the conversation, never copies you on your own reply, and leaves out no-reply and mailing addresses. A thread on another address of your connected account, such as an Outlook.com alias, counts as yours: Hive knows the addresses your mailbox sends from and never addresses the reply to any of them, and the reply goes out from your connected account. A reply-all to more than 25 people in To or in Cc is not prepared. Hive does not answer calendar invitations, notifications or mailings by email.
  • Hive remembers which of your email threads an answer read or replied in. When you follow up in the same chat (“reply in the same thread”, “use @Outlook to send it”), it replies in that thread without searching your mail again. Only your own follow-ups get this: in a shared chat, someone else’s question sees the answer as displayed, never the thread behind it. The same goes for the signal or matter a chat was opened from: a workspace admin who reads your chat, or anyone you share it with, sees the answers but not that context or those threads. “Let me approve first” gets you a card for the reply; “draft a reply” saves a draft in the thread instead.
Once an action has run, its card says what actually happened, on Draft after you approve and on Act when Hive acted by itself:
  • Saved as a draft — a draft was saved in your mailbox (or another tool), and nothing was sent. Before you approve, a card that only saves a draft (an Outlook or Gmail draft, a draft invoice) is marked Reversible: nothing goes out, and you can delete the draft.
  • Recorded in Hive or Draft created in Hive — an internal note or draft that reached nobody outside Hive. A workspace document reads Created in Hive or Updated in Hive. These never read as sent.
  • Only part of this ran — some steps went out and others did not. The card offers no Retry, because a retry could repeat the steps that did run.
  • Can’t confirm this went out — Hive cannot tell whether the action reached the provider. Check the provider before trying again; the card offers no Retry.
  • Couldn’t complete this action — nothing ran, with the reason. It may offer Retry. The reason names the app and what went wrong in plain words: the app refused the action (the account may be missing a permission, or the connection may need reconnecting), the app could not find the record, the app is limiting requests, the app was unavailable, the record changed since Hive read it, or the app rejected the action’s details (with the app’s own reason, where it gives one). Hive asks you to reconnect an app when the app no longer accepts the account’s sign-in, or, for Google apps, when the account is missing a permission the action needs: reconnecting is how you grant it.
A card that created a Google Drive folder links it: select Open in Google Drive. Later questions in the same chat know how each earlier card ended (done, declined, failed or still waiting) and, for a file or folder a card created, its link. Hive still reads the app again before it acts on that folder, so a card always shows what is there now.

Steps that build on each other

Some requests take two approvals in a row. When you ask Hive to move files into a folder that does not exist yet (“move the Q3 PDFs into a new folder Board packs”), it first shows a card that only creates the folder. Approve it in the chat and, once the folder is created, Hive goes on by itself: it sends Continue with the next step. for you and prepares the card that moves the files.
  • The next step never sends or changes anything by itself. Everything it prepares comes to you as a card, even through an account on Act and even in a chat with an agent, so the move waits for your approval.
  • Hive goes on only when every card on the answer is done. If you decline the folder card, it fails, or you approve it from Inbox rather than in the chat, Hive stops there: reply in the chat (for example “go on”) when you want the next card.
An answer only says an approval card is ready when Hive actually made one in that turn. When nothing was prepared (Hive needed more from you, a check refused the change, or the account is on Read), the answer says plainly that no approval card was prepared in that reply. When you ask for a change to something Hive just read in the same chat, such as a record it listed, it uses that read for the card rather than reading the app again. Actions you request in Ask appear in the conversation rather than on the Signals screen. Every completed action is recorded in the audit log.

See where each value came from

Each value on an action card has a line beneath it saying where Hive found it, so you can tell what you asked for from what Hive read, remembered or wrote itself before you approve: For a live read or a Brain memory, the exact read or fact is named on a second line underneath.
A value marked Brain memory, not read live was not checked against your connected tools for this action and may be out of date. Check it before you approve, especially an email address, link or recipient.
How Hive decides:
  • It looks for the whole value in your request, then your workspace’s recipient records, then what it read for this request, then the Business Brain, and shows the first place it finds it. A value that is in both a live read and Brain memory shows as read live.
  • Text Hive wrote, such as a message body, is marked as Brain memory when it quotes an email address or link that only Brain memory holds.
  • A row that shows several values reports Brain memory first if any of them came from it, and otherwise the least-checked source among them.
  • The line reports where a value could have come from, by matching its text. It does not trace how the model chose it.
  • Numbers, yes/no values, values under three characters and rows Hive fills in itself, such as the sending mailbox or a formatted amount, have no line. A few kinds of proposal show no lines at all, including proposals raised from signals and reviewed in the Review decision panel.
Hive stores each value’s source with the approval, next to the exact values you approve.

Hive acts on your own words

Hive plans an action from what you typed. It uses your last few messages in the conversation as well as this one, so a follow-up like “create a doc with the Q3 target and owner I just gave you” fills in those values. It never uses Hive’s own replies, or another member’s messages in a shared chat, as something you said, except a choice you pick from Hive’s list (below). To work out what a short reply such as “Yes, go ahead” refers to, it also restates your request, but it never treats that restatement as your words:
  • Contact details. When Hive creates or updates a record in a connected tool, an email address, phone number, website or profile link counts only if you typed it, in this message or an earlier one in the conversation, or Hive read it from a connected tool. One that appears only in Hive’s restatement or replies, or in another member’s message, is refused, and Hive asks you for it: “Give me the value and I will line it up.”
  • Pick from Hive’s list. When Hive answers your message by asking you to choose from a numbered or bulleted list, such as “Which Charlotte should I invite?” followed by two people and their email addresses, you can reply with just the number (“1”), an ordinal (“the second one”, “last”) or a word only one choice contains (“the Mastercard one”). Hive carries on with the choice you picked, and its details, such as the email address, count as yours, including in your next replies, and Hive can save them to your Business Brain without you typing them again. Only the choice you picked counts: never the line that introduces the list, the other choices, or a link’s hidden address. Hive reads your reply as a normal message when the list is not a choice it asked you to make (a list of findings, say), when the list answered another member’s message in a shared chat, when your reply matches more than one choice, is a question or says no (“not 2”, “the other one”), or when a bare number could be a choice’s own value (“3” after “3 months, 6 months, 12 months”).
  • Name the record. Before an irreversible action, such as deleting or archiving, when what Hive read holds several possible targets, you must name the one you mean, in this message or an earlier one. Otherwise Hive asks “Which company do you mean: A, B or C?” (up to five names) and prepares no approval. If the planned record is not the one you named, Hive refuses and says so. “Archive it” still works when Hive read only one candidate, and you can name several targets at once (“archive Acme and Globex”).
  • Drafts stay drafts. If your message mentions a draft anywhere (“draft”, “don’t send”, “do not send”, “for me to review”, “let me review it first”, “show me first”, “before sending”), Hive writes a draft, including a reply in the same email thread, and does not send it, even when your message or its restatement also says “send”. When it is unclear, Hive drafts. To send, ask to send without mentioning a draft or a review.
  • Sends stay sends. If your message starts by asking Hive to send (“send it”, “help me send email to each of them one by one”) and mentions no draft or review, every new email Hive writes for it is a send, so a sequence you asked to send does not switch to a draft partway through. Each still comes to you as a card to approve first.
  • One card per email. Ask for a separate email to each of several people (“draft an email to each address on this list”) and Hive prepares a card for each, up to 25 at a time, all through the same mailbox.
  • Long emails are never cut. A new email or draft can be up to 20,000 characters, and the card shows all of it. Hive never shortens an email to make it fit: a longer one is not prepared, and Hive says it is too long and offers to split it into several emails or put the full text in a document.
  • Your own mailbox. An email Hive writes goes through the Gmail or Outlook mailbox you have connected. If you have only one, Hive uses it. If you named a mailbox that is not connected, Hive says so and names the one you have. Hive says no mailbox is connected only when none is, and then the answer ends with Connect your email: choose Connect Gmail or Connect Outlook to connect one there and then. If your only mailbox needs reconnecting, the answer can offer Reconnect Gmail or Reconnect Outlook, with the mailbox’s address, instead (rolling out; see Connect or reconnect an app from the answer).
  • Suggestions from Hive’s monitoring are not your words. When Hive notices something in your connected tools and plans a fix, the description of what it noticed is never treated as something you asked for. A contact detail that appears only there, for example an address quoted from someone else’s email, is refused like any other unread value, so the suggestion asks for it instead of filling it in. Hive also never builds a ready-made email, a meeting invite or a HubSpot outreach from that description alone: the suggestion is planned like any other, with the description as context.
  • Links do not type for you. When a link opens Ask with text already in the box, such as Open in Ask on a signal or a matter, Hive leaves out any email address, phone number, web link, #channel or @name in that text. Type the ones you mean yourself: what you send counts as your own words, so text someone else put in a link cannot choose where Hive sends. A saved prompt’s Use in chat and the example questions in Hive’s first-run guide work the same way, except that an app pin such as @Outlook stays. Starting a chat from a project keeps everything you typed.

Recurring summaries, workflows, agents and things to keep an eye on

Ask for something on a schedule (“every Monday at 8, send me a summary of overdue invoices”), once at a set time (“remind me tomorrow at 9 to chase Acme”), a workflow, an agent, a run of one of your workflows, or something to keep an eye on (“tell me when the signed contract arrives”), and Hive shows a card with exactly what it would create or follow. Nothing is created or run until you press its button. See Build from chat and Ask Hive to keep an eye on something.

Use an agent in Ask

Choose an agent from the agent picker (Use default Hive is the default), or type @ and pick it from Agents, to answer with that agent’s instructions, knowledge and connectors. See Agent Studio. Both list only deployed agents you can run. An agent shared with you as Can view is not offered: ask its owner for Can run access. If your access changes to Can view while you set a chat up, Ask says You can view this agent but not run it. Ask its owner for Can run access. instead of asking for the agent’s sources. You can also run a deployed agent from its own page, which keeps its run history: see Run an agent directly. On a new chat, an agent with Conversation starters lists them directly under the box, one per row, in place of Hive’s suggested questions. This works whether you chose the agent in the picker, with @ or from a link that opens Ask with it. Select one to ask it. If the agent has a setup to complete first, or a required field is still empty, the starter goes into the box instead, so you can finish the setup and send it. The starters come from the agent version the chat uses. The Your data is ready card and the tools, such as Create document, follow below the starters; the tools still apply to the agent’s answer, and the card hides its own questions while an agent is chosen. An agent without starters keeps Hive’s suggested questions, and Use default Hive brings them back. Each answer the agent gives names it, with its icon, where a Hive answer says Hive: on the answer’s first line (for example Evidence reviewer · 12s) and while it works. Every answer keeps the name of the agent that gave it, so switching the chat to another agent, or back to Use default Hive, doesn’t relabel earlier answers, and answers Hive gave without an agent still say Hive. People you share the chat with see the same names and icons, even for an agent that isn’t shared with them. They see only who answered: the agent’s instructions, inputs and files stay hidden, and they can open or use the agent only if it is shared with them. An agent can do more when its author allows it:
  • Your data. An agent that declares The person’s own data (documents, Business Brain, Signals, Hive data or connected systems) reads those sources as you, with your own access. It never sees anything you could not read yourself, and never its author’s data unless the author is the one asking. An agent that declares none answers from its own knowledge only.
  • Actions. An agent can take only the actions its author selected for its attached apps under Integrations, for example Send email in Gmail or Create deal in HubSpot. It carries them out straight away through your own connected accounts, whatever each account is set to, within your workspace’s limits and the agent’s write budget, and the answer says what it did. An action whose recipient, channel, number or link did not come from the agent’s instructions or your own recent messages in this chat, such as an address from an email the agent read, waits for your approval as a card in the chat instead. Without an agent, Ask follows each account’s Read, Draft or Act level. See What agents and workflows do by themselves.
  • Other agents. An agent can hand part of the work to the agents its author lists under Sub-agents, if you may use them. They work as you, within the first agent’s budget, and their answers and sources come back into the first agent’s answer.
Tools and open files work with an agent too, within what the agent may read:
  • Business Brain. The agent answers from the Business Brain only when it declares the Business Brain under The person’s own data. Otherwise it tells you it cannot read it.
  • Chart. The agent charts a spreadsheet it found in its own knowledge, or in your documents when it declares them, or the numbers you typed.
  • Image. The agent draws the picture you asked for, if your workspace allows images.
  • An open file. The agent can read the file you have open in the chat. It changes the file only when you ask it to, and then only the part you asked about; see Change a file you have open.
Public agent links never get your data or actions. An agent built with the earlier connector settings, which named an account and records for each app, now answers in Ask too. It reads those apps through your own connected accounts, never its author’s, and is told which records it was set up for. Hive can’t hold it to those records, so it reads what your account can see. An agent whose connector settings include write operations, or one with a multi-step action plan or a connected template, still can’t be used in Ask. An agent with a run setup starts a chat of its own (below).

Agents with a run setup

An agent with a run setup (typed fields, or run sources such as a Library folder) works as a chat of its own:
1

Set it up

Pick the agent. Above the box, a small card asks for its inputs, sources and accounts. These stay fixed for the whole chat. If the agent needs nothing, skip straight to your message. A source starts filled in with the Library folder or file the chat already has selected, or else the folder of the project the chat is in, when the agent offers it to you ready. Nothing else is filled in for you: otherwise you choose the source yourself. Each source has one picker: click it to see what you can choose, or type in it to search. A chosen source shows as a chip in the picker; select × on the chip to remove it. Ask does not fill a cleared source in again. If the sources can’t be loaded, one line under the picker says so, with Retry.
2

Send your message

Type it in the box, which reads Message and the agent’s name, and press Enter. Your message appears as a bubble. While the agent works, the chat shows that it is working, under the agent’s name and icon, and Stop cancels the run.
3

Read the answer and follow up

The answer appears in the same card as a Hive answer, with Copy and the thumbs to rate it. Its first line names the agent, not Hive, beside the agent’s icon, with how long it took (for example Evidence reviewer · 32s); point at a long name to read all of it. When it used sources, made files or returned data, N sources · N files · Data below the answer opens them. Send a follow-up from the same box, if the agent allows follow-ups. A follow-up builds on up to 50 earlier answers; after that, select New chat.
After the first message the card becomes Inputs fixed for this chat · New chat. To change the inputs, select New chat, which starts a fresh chat with the same agent. The chat is listed in the sidebar like any other and opens as the same conversation after a reload. The agent reads only what its setup names. It does not use the chat’s project, Library folder, attached files, Google Drive or OneDrive files, connected-app pins, Tools or model choice, so those controls are hidden while it is selected. Picked inside another chat, the agent starts a new private chat, and the box says so. The agent’s chat can’t be shared. In a chat that is shared with you, you can still switch the box to Comment while the agent is selected. It reads with its own tools: its knowledge and run sources, the member data it declares under The person’s own data and its attached apps, read as you, and the web when its author allows it. It carries out the actions its author selected through your own connected accounts as each account’s level allows (straight away on Act, after your approval on Draft, never on Read), within your workspace’s limits, up to five a run, and records them on the run. One whose recipient, channel, number or link did not come from the agent’s instructions or what you typed, in your message or on its setup card, waits for your approval as a card instead. See What agents and workflows do by themselves.

Conversation memory

Within a conversation Hive remembers earlier turns, as long as the workspace allows chat memory (Settings → Products → Chat memory, on by default). Separately, you can let Hive use your own saved notes about you — see Personal memory.

Feedback, sharing and history

  • Rate an answer with a thumbs up or down. After rating, a box offers an optional comment of up to 2,000 characters. Hive saves your rating and comment and sends them to the Hive team, who read every one. The Hive team sees the rating, your comment, your workspace’s name, your role, the model that answered and when you rated it. It does not see your name, your email address or the chat’s questions and answers. In a chat with an agent, your comment also goes to the agent’s creators (see Feedback). Rating the same answer again replaces your earlier rating and keeps your comment.
  • To tell the Hive team how Hive is working for you in general, select Feedback at the top of Ask, choose Bad, Fine or Good, and add a comment if you like. It goes to the Hive team in the same way.
  • The whole conversation stays in the chat: scroll up to read every earlier question and answer. The message box stays at the bottom of the chat while you scroll, and Latest, just above it, takes you back to the newest message.
  • Your chats are listed in the sidebar, five at a time with Load more; group them into projects. On a computer, a chat’s or project’s ⋯ menu appears when you point at the row or tab to it; on a phone, tablet or touch screen it is always shown.
  • After the first answer in a new chat, Hive names the chat with a short headline of what it is about (at most six words), in place of your first question. This applies to chats in projects and chats with an agent too. For an agent’s answer or one built from selected Library items, the name comes from your question alone. Naming a chat is one small call on Hive’s lowest-cost model, charged to your seat’s credits like any other answer; on a BYOK seat it runs on your provider’s small model. If it cannot be done, the chat keeps your first question as its name. A name you choose with Rename is never replaced, and people you share the chat with see the same name. Chats from before this change keep their names.
  • While Hive is working on an answer in a chat, a small pulsing dot shows on that chat in the sidebar, in the place of its ⋯ menu. It shows for your own turns and, in a chat shared with you, for anyone’s. It stops pulsing (and stays as a still dot) if your device is set to reduce motion.
  • Deleting a chat removes its messages, files, documents and comments, and Hive then permanently deletes the stored files, as described in Attach files. If you used an agent in it, the agent’s runs and their usage stay in your workspace’s records, no longer linked to the chat. Those runs stay in the agent’s run history, marked No longer in its chat, read only: you can still open each one and read its answer, but no one can continue it. Your workspace’s admins see each run’s status and who ran it, but only the person who ran it can read its answer. The same happens if you share the chat or switch it to another agent.
  • An answer keeps generating if you leave the page. Come back or reload and you see the answer as far as Hive has written it, updated about every second until the checked answer replaces it. Only you see an answer while it is being written: people you share the chat with see the steps, then the finished answer. An answer from an agent or from selected Library items, or one you ask in a chat someone shared with you, appears only once it is finished, because Hive first checks what it may show. When Hive is updated while it is writing, it lets the answer finish first. Most answers you ask in the app run on Hive’s background service, and an update waits for them for up to 30 minutes, the longest an answer can run. An answer Hive had to write elsewhere, for example while that service was busy, gets about four minutes. One that cannot finish in its time shows “This answer was interrupted by an update.” with Retry instead of generating forever. See Troubleshooting.
  • As you scroll through a conversation, its header stays visible so you can return to chat controls without scrolling back to the top. The header is one row: choose the chat’s name to Rename, Save as agent, Share chat or Delete it (each offered only when you can do it). Save as agent refuses a draft that names an address, number, link, channel or repository you did not type in the chat: see Save a chat as an agent. The paperclip with a number lists the chat’s files: the files sent in the chat and the documents, sheets and decks made in it.
  • You can share a chat with colleagues in your workspace, if your workspace allows chat sharing. Sharing a chat shares everything in it, including answers Hive built from your apps, files and Brain. Share only with people who should see that.
    • Everyone you share it with, as Can view or Can collaborate or as a member of its project, sees every answer in full, with its citations and steps, even one built from a Library file, a connected app or an agent they can’t open themselves. In your own chat you likewise see in full every answer a collaborator asked. AI tools connected through the Hive MCP server can’t read Ask chats, shared or not.
    • They can open and download the files sent in the chat from Chat files (the paperclip at the top of the chat), read-only: the copy attached to this chat, never the Library or drive item it came from, and never files someone has added but not yet sent. Private attachment unavailable shows only where the chat’s copy of a file is no longer stored.
    • A Can collaborate member can attach their own files to their questions. Those files stay theirs: Hive reads them for that member’s questions and follow-ups, never for yours, and once sent, you and everyone else in the chat can open them from Chat files, read-only. Only the member who added a file can remove it.
    • The chat’s files (the documents, sheets and decks Hive made or anyone added in the chat) are yours, stored with your chat, whoever made them. Everyone you share the chat with can open them, see their history and download the original of an uploaded file. Can collaborate members can also edit them, ask Hive to change them and make new ones, and the history names who made each change. Can view members can only read them. Nobody but you can delete a file (deleting the chat removes its files), save one to a drive, or save one to the Library for the first time. To give someone one file and not the chat, use Share and export → Share this file in the file panel; see Share one file without its chat.
    • When a Can collaborate member saves a chat file that is linked to your Library, Hive writes their change to that one Library file as a new version, as you and only while you can still edit it; the file’s History in the chat names them, and so does the Library file’s Version history for you. If the chat has Library folders selected, their new files are filed there as yours, the way your own are. Either way they get no access to your Library: they can’t open, find, list or delete anything in it, that file included. A Library file you open into a shared chat shares that version: they can read it and download it as it was when you opened it, but get no access to the Library file itself or to its later versions.
    • When a Can collaborate member asks Hive something in your chat, Hive can make, edit and copy the chat’s files for them, and the files stay yours. When they ask, Hive can also open a Word, Excel or PowerPoint file you sent in the chat in the editor as a new chat file of yours; the file you sent isn’t changed. For them, Hive doesn’t save files to your Library or draw new pictures in your chat. It can’t put a Library template you can’t use on your files: it says “This template isn’t available to the chat owner.”
    • A Can collaborate member can also brief Hive AI on the chat’s decks in the deck editor, with their own credits and access; the request and answer are saved in your chat. Can view members can’t. Someone a deck is shared with on its own as Can edit briefs Hive AI in a private chat of their own, never in yours. See Who can brief Hive AI on a deck.
    • A Can collaborate member with the right role can also use the Workflow & schedule builder in your chat. Whatever they create, schedule or run is theirs and runs as them, never in your account, and only they can press the cards from their questions.
    • If the chat is filed in a project, the project’s instructions and skills apply to everyone’s questions in it, including someone you shared the chat with directly who isn’t in the project. The project’s pinned files stay with its members.
    • The Library files, apps and agents behind an answer stay as closed to them as before.
    • When someone asks a follow-up, Hive remembers each earlier answer they can see as it is shown, with its text and citation titles. It doesn’t re-read the Library files, agents or attached files behind it for them, and it never reads your attached files, or the Library files you picked for the chat, to answer their question (opening a sent file in the editor when they ask is the one exception). Your Library pick isn’t shown to them as the chat’s Library context.
    • If you later lose access to a Library file or agent behind one of your own answers, that answer shows as unavailable to you, as in a chat you never shared.
    • Answers an agent gave show that agent’s name and icon to them too, even if the agent isn’t shared with them; see Use an agent in Ask.
  • In a chat someone shared with you, the header shows one line such as Shared by Mir · Can collaborate. Choose the info icon next to it to see how you have access and since when.
  • Under the composer of a shared chat, Hive reminds you: “Everyone in this chat sees Hive’s answers. Hive reads your connected apps as you.” Each question reads the connected apps of the person who asked it, never the chat owner’s.
  • A shared chat answers one question at a time, and a question asked meanwhile waits its turn. While Hive answers someone else, the composer says who asked, for example “Mir asked Hive. It’s answering now. Ask now and your question runs after it.” You can post a comment straight away too. Only the person who asked can Stop that answer, and only the author of the latest question can use Edit to change and resend it.
    • A question you ask then is queued: it shows under the chat as Queued, runs after the current answer, with who asked it, and everyone in the chat sees it.
    • In the deck editor, Hive AI works on one request at a time in a deck, and a follow-up you queue there runs after it; see Follow and review Hive AI’s work.
    • Hive starts it by itself as soon as the answer before it ends, whether that answer finished, failed or was stopped, even if you have closed the chat. Several waiting questions run one at a time, in the order they were asked, and once questions are waiting, a new question joins the end of the queue rather than going first.
    • It runs as you, exactly as if you had asked it then: with your access to apps and files, on your seat and your credits, and named Asked by you. Hive checks again that you can still ask in the chat when it starts.
    • Until it starts, you can Edit the words of your own queued question (it keeps the tools, mode and model you asked it with) or Cancel it. Others see it but can’t change it.
    • Up to 5 questions can wait in a chat. Beyond that the composer says “5 questions are already waiting: ask once one has run.”
    • A queued question can’t carry files of its own: remove them, or ask once the answer finishes. It can be for an agent you picked while Hive was answering. Hive checks that you can still use that agent when the question starts.
    • If you lose access to the chat, or leave the workspace, before your question starts, it isn’t asked, and a note under the chat says why. The same happens if Hive can’t start it as you, for example because you have no seat or no credits left; only you see the reason. If Hive can’t start it because of a passing problem on its side, it tries again a few minutes later, up to three times. The next question then runs. You can dismiss a note on your own question; it goes by itself after a day.
  • When someone asks a question in a shared chat, or Hive finishes, fails or stops an answer, every open copy of the chat updates by itself within a few seconds: yours in another tab, and your teammates’. You don’t need to reload.
  • Each Hive answer names who started it, for example Asked by Mir. An answer that no question asked for says what posted it instead: Continued after your approval decision when you approved or declined a card and Hive answered on, or Sent from Save to Drive by you for the approval card Save to Drive posts. Every question stays in the chat, including one that was only attached files or @-pins.
  • In a shared chat, collaborator comments appear in the same timeline as questions and Hive answers, in the order they were posted. They load alongside the chat and take their place between the messages as soon as they arrive. A long chat loads all of its comments, so the newest are always there without paging. Each comment keeps its author’s name; your own comments are labelled You. Use the edit or delete icon on a comment you wrote to change or remove it. Your own questions and comments sit on the right; your teammates’ sit on the left, as do Hive’s answers. Questions asked before Hive recorded who asked them were the chat owner’s, so they sit on the owner’s side and carry the owner’s name.
  • Comments can’t take attachments at present. In Comment mode the composer has no + and doesn’t accept dropped files; switch back to Chat to attach files to the conversation.
  • The composer of a shared chat switches between Chat (ask Hive) and Comment (write to the people in the chat). If you switch after typing, your text comes with you; an app or person you picked with @ comes across as plain text, so it no longer pins the app or mentions the person. If the other mode already holds a draft of its own, Hive shows that one instead and keeps yours for when you switch back.
  • Type @ in a comment to mention anyone in your workspace who holds a Basic, Pro or Max seat (a member without a seat can do nothing in the workspace, and a member on an Approver seat sees only what is routed to them, so neither is offered). Someone the chat isn’t shared with is marked Not in this chat. When you post, Hive asks whether to share the chat with them first:
    • Share and send shares the chat with them as Can collaborate, so they can read and reply to your comment, then posts it. They’re notified about the mention.
    • Send without sharing posts the comment, but they aren’t notified and the comment says they can’t see it.
    • Cancel keeps your draft and posts nothing.
    Only the chat’s owner can share it, so other members see why the share option is missing and can still post or cancel. Being a workspace admin doesn’t make you part of a chat: an admin is asked about in the same way until the chat is shared with them.
  • When someone mentions you in a comment, the chat gets a small red dot in the sidebar, and so does its project if it’s in one you can see. Screen readers announce it as “Unread mention”. The dot is in the phone menu too. It goes when you open the chat, or when you mark the mention read in the bell or the Inbox. Only you see dots for your own mentions, and a chat you can no longer open shows none.
  • Type # in a comment, or select the paperclip in the comment box, to refer to a Library file. The list shows only Library files you can open; type to search it. The file appears in your comment as a file chip:
    • Anyone in the chat who can open that file sees its name, and selecting it opens the file in the Library.
    • Anyone else sees A Library file you don’t have access to, with no name. A file that’s deleted later shows the same way for everyone.
    Hive checks who can open the file each time someone reads the comment, so if the file is shared with a reader later, the chip turns into a link for them. Referring to a file doesn’t share it with anyone or copy any of its content into the comment. A comment can refer to up to 10 Library files. If you edit a comment that refers to a file you can no longer open, the reference stays for the people who still can.

Ask from Slack or Teams

If your workspace has installed a channel, mention @Hive or send it a direct message to ask from Slack or Microsoft Teams. Answers there are read-only: signals and approvals stay in Hive. An answer that cites sources ends with a Sources line that names each origin once, with the same labels as the web (see Citations). See Slack and Teams.

Tools and modes

Documents, decks, spreadsheets, charts, images and web search.

Deck editor

Edit, share, present and download a deck, with Hive AI built in.

Projects

Shared instructions and files for a group of chats.

Models

Choose which model answers, and how admins restrict the list.

Business Brain

What Hive knows about your business.