Skip to main content
Hive uses large language models to read, reason, write and plan. This page explains which providers are involved, what data they receive, and the controls you have.

Where AI is used

In every case the model proposes. Whether anything happens in your tools is decided by Hive’s deterministic policy engine and, where required, a person. See Security overview.

Model providers

Hive works with models from several providers:
  • OpenAI
  • Anthropic
  • Google
  • xAI
  • Microsoft
Hive sends requests to these model providers to answer you, using the model you picked. Each request carries what that answer needs, including the excerpts from your tools described below. Embeddings (used to search your knowledge) and image generation use OpenAI models. With your own key, members’ model work runs on your key; Signals, background work, images and knowledge search still run on Hive’s accounts. In selected workspaces, Hive’s own decision model for Signals, run in Hive’s own AWS account in Frankfurt, decides whether a matter needs you, what it needs and how urgent it is: see The Signals decision model. Web search in Ask, in agents and in workflow agent steps uses Tavily, a search provider; a workflow’s Web search step runs on OpenAI’s hosted web search instead. Dictation (voice input) is transcribed by Groq (Whisper Large v3 Turbo): Hive’s servers send each spoken phrase, keep no audio, and record only how long it was. Which models appear in your model picker depends on what is configured for your workspace; Hive does not guarantee a fixed list. Hive checks every hour which models are currently available and stops offering any that no longer are; a default or saved agent set up on one runs on the next available model of the same provider and tier, without your settings being changed (see Models that are retired). Each answer records the model that actually produced it, and Hive records the model that actually served every model call, with what it used, so your workspace’s usage and credits follow the real model even when a request was routed or retried on another one.

What providers receive

To answer a request, the model needs context. Hive sends only what that request needs:
  • your question or the step’s instructions;
  • the relevant excerpts retrieved from your tools, files and Business Brain;
  • earlier turns of the conversation when chat memory is allowed.
Hive does not send connector credentials, API keys or secrets to any model. Untrusted content — emails, documents, web pages — is clearly marked as data, never instructions. That marking names only the kind of source, never your workspace, account or file IDs. Hive removes it wherever a model quotes it back: in an Ask answer while it streams, in an agent’s or workflow’s written result, in approval cards and meeting recaps, and in a reply Hive prepares for you to send. Documents and agent results saved while the marking still named your workspace are shown without it; what is stored is not rewritten. The marking covers more than tools’ results:
  • Workflow run data. The values a run fills into an AI step’s prompt — form answers, an app event, earlier steps’ answers, web results — are marked as data; only text the workflow’s author wrote, and what a member typed into a run they started in Hive, is not. A workflow’s automatic Web search sends the run’s data as a topic only: it never opens, or searches for, a web address written in it or built from it, and opens only pages its own searches return.
  • Lookups. When Hive looks something up in a connected app, the request it reads from is marked as data too, since it may be written from other text.
  • Agents’ attached files. An agent’s attached files, documents and project knowledge reach it at the start of the conversation, marked as data; its attached skills stay part of its instructions. A PDF or image sent to the model whole comes with Hive’s note that it is data, under a neutral name (“Attached file 1”), so a file’s own title cannot pose as an instruction.
  • Agent Studio’s Build chat. The names and descriptions of the workspace’s agents, skills, files and tools it is shown, and its own earlier replies, are marked as data.
  • Signals. What Hive wrote from your mail — a matter’s title, next step and summary, and names in your Brain — reaches Signals’ reasoning quoted, never as Hive’s own notes, so a sender cannot pass off a correction or a mute by writing one into an email.
Hive does not send prompts or completions to third-party tracing or observability services.

What Hive records about each model call

To keep AI fast, reliable and affordable, Hive sends a short record of each model call to its product analytics provider, PostHog (EU). The record holds metadata only:
  • the model and its provider (a model you added with your own key is shown only as custom:<provider>, never by name);
  • which part of Hive made the call, such as Ask, a workflow run or Signals;
  • token counts: input, output, cached input and cache writes;
  • how long the call took, and Hive’s own estimate of what it cost;
  • whether it failed, as a fixed code such as timeout or model_refusal, and why the model stopped;
  • for an answer that came back in the wrong shape, the names of the expected fields that were missing or invalid;
  • a trace id, so the calls made for one request can be read together. It is a one-way hash of the workspace and an internal key, so it carries no member id or text you typed.
What happens to the record depends on who started the work:
  • Work a person started, such as an Ask question, a workflow or agent run started by hand, or a question in Slack or Teams, is sent call by call, grouped by request.
  • Background work, such as scheduled runs and Signals checks, is mostly sent as 15-minute totals per workspace, model and feature: call counts, tokens and cost. A small sample of background calls, and some failed ones, are sent call by call so problems can be traced.
  • Embeddings (used to search your knowledge) are only ever sent as totals.
Records are kept at workspace level. A record is linked to you only when you made the request in the web app and accepted analytics; see Product analytics. No record carries a prompt, a completion, a tool’s input or output, a file or account name, or the text of an error.

What can leave Hive after a private read

Once an Ask answer has read your private data — your files, connected apps or workspace records — Hive’s privacy rule limits what that answer may send to other outside services. An agent’s run counts as having read private data from its first step, because its instructions carry its knowledge:
  • Web searches and web pages are never held back. When you ask Hive to search, it searches, whatever the answer has read: the query is written from the whole conversation, and that includes what it read, which the search provider receives. Hive still opens only pages a search returned, links you wrote and same-site links on pages it opened, and treats what it reads there as data, never instructions. See Reading whole web pages. A server that searches or reads the web — Exa, Tavily, Firecrawl and Perplexity — and GitHub’s searches across all of GitHub count as web searches here too. The exception is Exa’s page fetch and crawl (web_fetch_exa and crawling_exa), which open whatever address they are given: once an answer has read private data, they are not called again in that answer. An Exa search’s results are a read from one of your connected apps, so that includes any answer that has already searched Exa, and every agent run. Exa’s searches still run.
  • Public register searches run after a private read too. This is rolling out to workspaces gradually. Once it reaches yours, Hive’s own searches of official public registers — Companies House (companies, officers and disqualified directors), The Gazette, the Charity Commission, the FCA Register, Brønnøysund and the other registers in the connector catalog — run after a private read, like a web search. As with a web search, the search text is written from the whole conversation, so it can include what the answer read, and the register receives it. Hive treats what a register returns as data, never instructions. A client onboarding agent with your client register in its knowledge can then check the client’s officers against the disqualified directors register and The Gazette in the same run. Searches of GOV.UK’s HMRC guidance and the ONS take any words and are not registers, so they stay under the rule for other outside lookups below. Until the change reaches your workspace, register searches follow that rule too.
  • Other outside lookups may use only your own words from the chat, your company’s public name and website, and plain search words such as “reviews” or “news”.
  • MCP servers you connected are your connected apps under this rule, like Gmail or HubSpot, whether they are in Hive’s directory or added by address: Ask and your agents keep reading them after a private read, through your own connection. Changes still go through an approval card or an agent’s selected actions. A server’s read that reaches outside your own data but is not a web search — such as Daloopa’s or Morningstar’s market data — is not called again in that answer. See MCP servers.
  • Pictures are the exception. A picture’s description goes only to Hive’s configured image model, which today is OpenAI’s, whatever model your chat runs on. OpenAI processes it as a model provider, like your chat’s own provider, and it may be a different provider from the one your chat uses. A description may use anything from the files you can read, such as the brand colours of a pinned project deck. The image request carries only the description: Hive never opens a link in it, and it goes nowhere else. The picture is stored in your chat, as the answer is.
  • Workflows’ automatic searches and pictures leave your apps out. A workflow’s Web search or Generate image step that builds its query or prompt automatically leaves out what came from steps that read your connected apps: an agent step with Read-only data sources, an Agent library step, a Read from an app step, and anything built from them. If nothing else is left, the step fails with “This web search step’s only run data came from steps that read your connected apps, which an automatic search leaves out. Give it a fixed query, or turn on web search in that agent step.” (a Generate image step fails with “This image step’s only run data came from steps that read your connected apps, which an automatic prompt leaves out. Give it a manual prompt.”), and deploying the workflow warns you first. See Web search and Generate image.
Owners and Admins can turn web search off for the workspace under Settings → Chat capabilities. That removes web search and web page reading from Ask and from agents. It does not turn off a workflow’s Web search step or web search in a workflow’s Hive agent step, and it does not remove an MCP server you connected that searches or reads the web, such as Exa, Tavily, Firecrawl or Perplexity: remove the step, or disconnect the server. When the rule holds a step back, nothing is sent, and the step shows Held back by Hive’s privacy rule: nothing was sent. Hive is told it was its own rule, not a failure of the tool or its provider, and later messages in the chat know which earlier steps were held back or failed.

Training and your data

  • Hive does not use your workspace’s chat and workflow data to improve its product unless an Owner or Admin turns on Help improve Hive in Settings → Workspace → Data sharing. With it on, the product states: “No personal data will be used.” Content Hive gets from Google services (Gmail, Google Drive, Docs, Sheets, Slides and Calendar) is never included, whatever the setting, in line with the Google API Services User Data Policy, including its Limited Use requirements. See Data protection.
  • Content from Google (Gmail, Google Drive, Docs, Sheets, Slides, Calendar and Analytics) is not used to develop, train or improve AI models: it is excluded from Help improve Hive, and Hive’s own Signals decision model is trained only on invented example mail. See Google user data.
  • Model providers process requests under their own terms for API and enterprise use. If your organisation has specific requirements about a provider, use the model restrictions or bring your own key, below, and contact us with questions.
  • Automated quality checks on Signals. Each night an AI model scores the signals Hive raised the day before, to check how useful they were. It reads only what Hive itself generated: the signal, its suggestion or draft, and the matter it belongs to. It never reads your emails, files or connected apps. The check respects your model settings: if your workspace’s model allow-list does not include the model it uses, or your workspace uses its own provider key, your signals are not checked. No person reads the signals for this: Hive’s team sees only counts, scores and reason codes for your workspace. Scores are kept for 90 days and are never used to train models. Hive pays for these checks; they never use your workspace’s credits.

Control which models are used

Owners and Admins can restrict the workspace to an allow-list of models in Settings → Workspace → Models (up to 50 models). An empty allow-list means no restriction.
  • Members see only models that are available to them.
  • Admins also see unavailable models, with the reason — for example, the provider is not configured.
  • The picker shows Workspace default, or Hive Auto when automatic routing is the default.
  • Hive Auto’s Fast model, which answers most questions, must be an Efficient model (Claude Haiku included), or a custom or unpriced model that only your organisation’s own key runs. A Balanced or Frontier model saved as Fast earlier is not used: Hive Auto answers on the workspace’s Efficient model instead. See Hive Auto’s Fast model.
  • An explicit model choice also governs the planning and retrieval steps behind that answer, including the steps that choose and summarise reads from connected apps.
  • Background work with no member’s choice behind it, such as Signals and Brain upkeep, runs on an Efficient model: the workspace’s Default background model, which may only be an Efficient model other than Claude Haiku, or Hive’s Efficient default. It is included in each Max seat.
  • Work started by a schedule, a connected app event or a signal runs on Efficient or Balanced models on Hive’s keys; a Frontier model chosen for it is refused. Public agent links and public forms run on an Efficient model unless the publisher turns on Balanced under Model for visitors. Frontier models never run there: if a public link’s agent or a public form’s step uses a dearer model than the publisher allows, Hive answers on its Efficient model instead.
  • Custom models and -pro models never run on Hive’s keys: they run only on your organisation’s own key. Custom models (models your workspace added itself) come with Enterprise.
See Models for the member view.

Bring your own key

On Business, Team and Enterprise, if your organisation has its own agreement with OpenAI or Anthropic, an Owner or Admin can store the organisation’s key in the Keys section of Settings → Workspace → Models. Members’ model work then runs on that key and is billed by your provider under your agreement. One organisation-wide key is used, and only its last four characters are ever shown. Adding, replacing or removing the key is recorded in the Audit log with the provider and who made the change, never the key itself. On Business, model calls on your key use none of your AI credits, and Hive adds no fee. On an earlier plan, Hive still meters its own part of the work: each model call it orchestrates on your key uses 1 credit from your workspace’s pool (an agent step that calls the model once per tool it uses, then once to answer, uses 1 credit for each of those calls). Images, web search (including workflow web search steps), knowledge search and the other services Hive runs on its own accounts always run there, whatever provider your key is for, and use credits as usual. Your seat prices do not change. See Billing and plans. Signals and Hive’s other background work run on Hive’s account and Hive’s Efficient model, not your organisation’s key (see Signals preparation below). Every other model call goes to your key’s provider:
  • Hive’s own defaults follow your provider. Steps that use Hive’s built-in default model — summaries and other built-in features, and Ask when your workspace sets no default model or model allow-list — run on your provider’s equivalent model: Claude Sonnet or Claude Haiku on an Anthropic key, GPT-5.5 or GPT-5.4 Nano on an OpenAI key. So do the Efficient models Hive chooses for you: quick answers and the Efficient model visitors to a public agent link or form are served.
  • Quick answers keep working. When the pool can’t pay the 1-credit fee, a member’s limit is reached or a payment is overdue, each Pro or Max member’s quick answers (10 a day and 200 a month) still run on your key, and Hive waives their fee.
  • Visitors keep to the public budget and the model for visitors. A public agent link or form on your key charges its 1-credit fee to the public budget, never to the member who published it, and uses top-ups only if an Owner or Admin allows public traffic to. Its Model for visitors applies on your key too: a Frontier, -pro or custom model runs as Efficient for visitors.
  • Defaults your workspace sets are kept. A default model or allow-list an admin sets in Settings → Workspace → Models, and a member’s own default, count as a choice: set them to models your key’s provider runs.
  • A model you choose is never swapped for another provider’s. Choosing a model from a different provider than your key is refused with a message saying so, so pick one your key can run. When a model fails before any of the answer appears, Ask tries once more on another model your key’s provider runs, never on Hive’s key (see When a model fails).
  • Workflows run on your key too. Workflow agent steps and AI conditions follow the same rules as Ask. Web search and image steps run on Hive’s account and use credits, as pictures in Ask do.
  • Custom and -pro models run only on your key, never on Hive’s.
  • A failing key stops the work. If your provider refuses the key, for example because it is rejected or out of quota, the work that runs on it stops and Owners and Admins are told: once a day while the same key keeps failing, and again as soon as a replacement key fails. Hive never falls back to its own key.
On a plan without your own key. The trial and Starter do not include your own key. A key stored on such a plan, including one stored before the workspace moved to it, is not used: members’ model work runs on Hive’s models and Hive’s provider accounts, under Hive’s terms with those providers, and uses credits from the pool as usual, and agents set to a custom model do not run. Settings → Workspace → Models says so under Keys, and offers no way to add a key, though you can still remove one stored earlier. If your organisation’s work must stay on its own provider agreement, keep the workspace on Team or Enterprise, or contact us before it moves.

Signals preparation

When Hive prepares a response to a signal — on its own, or because you chose Act on this — it uses your workspace’s Default background model from Settings → Workspace → Models, which is always an Efficient model, or Hive’s Efficient default when none is set. When Hive prepares the same signal again, it uses the model that first prepared it, as long as your workspace still allows that model. If it doesn’t, Hive uses the current default. This applies to every retry and to Act on this. Every step of a preparation runs on that one model: reading what the signal is about, writing the plan, checking that each proposed action is backed by what it read, and writing the signal’s card — the response for you to approve, or what Hive found or did. On Business, Signals preparation uses the AI credits of the member whose accounts it reads, at cost. On an earlier plan, Signals preparation is included in each Max seat, up to the seat’s fair use of prepared actions a month (25 on Starter, 120 on Team and Enterprise). Above it, a prepared action uses about 11 credits from the pool (a client reply about 12), up to each member’s limit for Signals above fair use. See Fair use and credits. Hive records the model and token usage of each preparation. With your own model key, Signals preparation still runs on Hive’s account, not your organisation’s key. Hive’s own background reading is included in the same way: its hourly check of your connected tools, the first sync after you connect an app, sorting incoming app events, setting up what to monitor in a connected app and extracting meeting recaps.

Replies are checked against what Hive read

Before a reply or message Hive wrote can go out on its own, Hive checks it against what it actually read in that run: the email thread, the signal’s own details, your Brain’s facts about the people involved, and live reads of your connected systems. Every amount, long number (such as a phone or account number), link and email address in the reply must appear, whole, in what it read.
  • If one does not, the reply is not sent. It waits for you as an approval card, even when the account is set to Act, and the card says why, for example “This reply mentions £5,000, which Hive could not find in the thread or anything else it read, so it waits for you.” The line never repeats a link or an address.
  • Hidden characters in a reply hold it too, except emoji joiners and direction marks the thread itself uses.
  • Hive never replaces the reply with a canned one, rewrites it or drops it.
  • Proposed days and times are not checked, so a reply suggesting a new meeting time still follows your account’s setting.
Signals sends an email on its own only as a reply inside your own thread on an Act account, and the check is stricter there: Hive must have read nothing beyond that thread to write it (no other thread, matter, connected app, Brain or knowledge search, or MCP tool), and every amount, long number, link and address in the reply must appear in that thread or the signal’s own details. Otherwise the reply waits for you as a card, for example “Act account: this reply waits for you because Hive read more than this thread to write it.” Every other email Hive writes for you in Ask or Signals comes to you as a card first. See Email replies on an Act account.

The Signals decision model

Hive has its own small decision model for Signals. It answers fixed questions about each matter with a score for each answer:
  • does this matter need you;
  • what does it need (a reply, an invoice to chase, a renewal and so on);
  • how urgent is it;
  • which of your matters do the new items belong to.
Hive turns it on question by question, for selected workspaces. When it is on for your workspace:
  • It decides; the Signals model still writes. The Signals model still reads new items, groups them into matters and writes every title, summary and card. For the questions that are turned on, the decision model’s answer replaces the Signals model’s: whether a matter gets a signal, its kind, its urgency, and which matter new items join.
    • Hive’s own rules still apply after it: card disputes, muted matters, and mail you were only copied on.
    • When the decision model raises a signal the Signals model did not, the card says “Hive’s decision model found that this matter needs you.”
    • The decision model still answers with its original kinds, so a finer kind the Signals model finds is kept when the decision model’s answer is its nearest original kind: a Deadline due on a reply owed, a document waiting or a renewal; a Payment due on a reply owed, a document waiting or an invoice to chase; a Follow-up due or Deal gone quiet on a reply owed. Any other answer replaces it, as for every kind.
  • No fallback. If the decision model is unavailable, Signals waits: the new items stay unread and are decided on a later hourly pass. Hive does not hand the decision to another model in the meantime, and it is alerted when this happens. A batch the decision model cannot answer at all (one it rejects, rather than an outage) is handled like any batch Signals cannot process: tried three times, then set aside until its items change.
  • Less input, to one place only. For each matter decision, Hive’s servers send the decision model only the new items the decision is about and the titles of your open matters they could belong to. It receives no Brain entries, no matter summaries or corrections, and no connector credentials. It runs on a server in Hive’s own Amazon Web Services account in Frankfurt (eu-central-1). The server is reached only over Hive’s encrypted private network, with nothing on it open to the internet: Hive sends this input only to a private network address.
  • The model server keeps no text. It runs with request logging off and answers from memory. It writes no prompt or input text to any log or file.
  • Only IDs and scores are kept. Hive keeps the IDs of the records involved, the decision model’s answers and scores, and timings, to measure the model and improve it. It keeps no text from that input, no names and no workspace name.
  • It is not trained on your data. The model is trained only on invented example mail written for that purpose, never on your workspace’s content. Hive tests it on real workspace data only in workspaces with Help improve Hive turned on, and never on content from Google services.
  • Fair use is counted once. The Signals model still reads every pass. When a pass is retried after an outage, the same items are not counted against fair use again.

When credits run out

On Business, each member has their own AI credits: when a member has used them, their AI work stops, unless an Owner or Admin has turned on extra usage, which carries on, billed at cost, up to the member’s top-up limit and the workspace’s monthly maximum. Background work on their accounts waits for a later pass, and scheduled runs wait up to 72 hours. See When credits run out. The rest of this section describes earlier plans. When your workspace’s pool and top-ups run out, work that uses credits stops until an Owner or Admin adds credits. The exception is postpaid overage: annual Team workspaces can turn it on, and Enterprise contracts include it. With overage on, work keeps running up to the overage ceiling and is billed at list price. When a member reaches their own monthly credit limit, their work stops until an Owner or Admin raises that limit in Billing or the next usage month starts. Adding credits does not lift a member’s limit. Each member on a Pro or Max seat keeps 10 quick answers a day and 200 a month, short answers on an Efficient model with no web search or files: a message that qualifies runs on that model, whatever model you picked. Apart from those quick answers, Hive never switches a conversation to a different model to keep it going when credits run out. (A model that fails before any of an Ask answer appears is tried once on another model your workspace allows: see When a model fails.) Scheduled and event-started runs wait up to 72 hours for credits to return. An agent’s When limits are reached setting doesn’t change this: Switch model only moves a run to Claude Haiku when the agent’s own model is rate-limited, overloaded or unavailable, and a run whose models can’t take a step moves once to another model your workspace allows (see Models for agents). Running out of credits always blocks the agent rather than switching its model. Signals within fair use and connecting an app are not affected: Signals’ work inside each Max seat, including the first read of a newly connected account within the seat’s included first reads, and the hourly check of your connected tools don’t draw on the pool. A first read beyond those uses credits at measured cost, and Hive says so before you connect the account. Above fair use, Signals keeps analysing items linked to your open matters and priority senders at Hive’s cost, and other items arrive without a prepared draft. See Fair use and credits.

Keeping AI output honest

  • Citations are built by Hive, not the model. The model points at evidence; Hive fills in the labels and links from the actual sources, and only allows safe https links.
  • Links show where they go. A link in a Hive answer in the web app shows its words and then its address in brackets, and only the address is a link, which opens in a new tab, so link text cannot disguise where it goes. A bare link shows once. Answers in Slack and Teams show the address too: see Slack and Teams.
  • Live-read failures are disclosed. If Hive could not check a live source, the answer says so and notes that it relied on stored data that may be out of date.
  • Confidence is labelled. A signal shows whether its confidence is Hive’s own estimate or measured.
  • Low-trust sources stay low-trust. Knowledge derived from an untrusted source is never promoted to trusted by the model.

Models

Choosing a model in Ask.

Data protection

Retention, sharing and deletion.

Security overview

Policy engine, approvals and isolation.

Infrastructure

The providers Hive runs on.