> ## Documentation Index
> Fetch the complete documentation index at: https://docs.get-hive.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Signals: how Hive catches what is slipping
> What a Hive signal is, how Hive finds signals in your connected tools, how they are ranked, and the evidence behind every one.
A **signal** is something Hive found in your connected tools that is at risk of being missed: an invoice
that has gone overdue, a customer email nobody has answered, a deal that has stalled, a metric that has
moved well outside its usual range. Each signal explains what happened, why it matters, and what Hive
suggests doing next.
Signals live on the **Signals** page in the sidebar (`/signals`). The page describes itself as "your
attention center across connected tools. Hive finds risks, explains why they matter, and proposes the
next safe step."
## What a signal contains
Every signal is built from the same parts, so you can read any of them the same way:
| Part | What it tells you |
| - | - |
| **Type and entity** | What kind of situation this is (for example "Failed payment") and which customer, invoice, deal or account it is about. |
| **Severity** | **High**, **Medium** or **Low**. |
| **Why it matters** | A plain-language explanation and a short stake line: money at risk, a count, a deadline, or who is blocked. |
| **Readings** | Up to six pieces of evidence, each with a value, a comparison (for example "against your usual") and the connector and field it came from. |
| **Source link** | Where available, a link back to the record in the source tool. |
| **Next step** | What Hive proposes, and sometimes a **prepared action** such as a drafted reply. Nothing in a prepared action has been sent. |
| **Confidence** | Either "N% — Hive's own estimate" (the model's self-assessment) or "N% confidence, measured". Hive labels which is which. |
Hive never stores the label of the connected account (such as a mailbox address) on the signal itself; the
screen resolves which account a signal came from when you open it.
## How Hive finds signals
Hive watches your tools in several complementary ways.
```mermaid theme={null}
flowchart LR
A[Connected tools] -->|hourly collect,
webhooks, Run loop now| B[Readings]
B --> C{Detection}
C -->|metric breaks
its baseline| D[Candidate signal]
C -->|mail waiting
on a reply| D
D --> E{Alert gate}
E -->|urgent, actionable,
not a duplicate| F[Signals feed]
E -->|repeat or noise| G[Updated, grouped
or held back]
F --> H[You act, snooze
or dismiss]
H -->|feedback| E
```
### The always-on loop
Hive runs an **hourly loop** across every live connector: it collects fresh data, runs detection, and
reflects on what it has learned. Some tools also push changes to Hive as they happen through webhooks, and
an administrator can start a run immediately with **Run loop now** on the Signals page (the button also shows
when the next scheduled run is due).
### Baselines, not single spikes
For numeric metrics, Hive learns what "normal" looks like for your business using a robust baseline (the
median and spread of recent values), then measures how far a new value sits from it. A value only counts as a
breach when it is far outside the baseline on at least **two of the last three** readings, so a single odd data
point never raises a signal on its own. The further outside the baseline a value is, the higher its severity.
### A monitoring plan per connector
For each connected tool, Hive writes a **monitoring plan**: the specific metrics it will read and how it will
judge them. Each metric is either a **state** (is something true right now, such as "invoices overdue") or a
**trend** (is something moving unusually). Hive dry-runs every read before adding it to the plan, and keeps only
reads that return a real number. You can see, pause and edit the plan on the [Coverage](/signals/coverage) tab.
### Mail detectors
For Gmail and Outlook, Hive looks for conversations where **someone is waiting on you** (the last message came
from them) and conversations that have **gone quiet** when you are waiting on someone else. Mail signals carry
no address, subject or excerpt in the signal itself; they point you back to the thread.
### Connector-specific detectors
Hive also has detectors built for particular tools, such as invoice collection for Stripe and Xero, deal
lifecycle for HubSpot, and stale assigned work for GitHub and Jira. These are rolled out per workspace and are
not switched on everywhere by default. See the [integrations catalog](/integrations/catalog) for what each tool
supports.
## Severity and ranking
Severity is **High**, **Medium** or **Low**. Hive ranks the feed by:
1. **Stage** — what needs you comes first (see below).
2. **Severity** — high before medium before low.
3. **A priority score** that combines severity, confidence and blast radius — how much the affected customer,
deal or account is worth to the business, based on what the [Business Brain](/brain/business-brain) knows.
How recently and how often a signal has recurred can lift its score, but never becomes the leading factor.
## Stages
Every signal sits in one stage, shown as a label on the row:
| Stage | Label | Means |
| - | - | - |
| `needs_you` | **Needs you** | New, or Hive tried to prepare a response and could not. |
| `in_progress` | **Hive is on it** | Hive is planning or carrying out a response you asked for. |
| `watching` | **Watching** | Hive is observing in shadow mode and will not act. |
| `snoozed` | **Snoozed** | Someone chose to decide later; it comes back on its own. |
| `closed` | **Closed** | Handled, approved, denied, resolved on its own, or dismissed. Hidden from the feed by default. |
When a condition clears by itself (the invoice gets paid), Hive closes the signal for you.
## Noise control: gating, grouping and fatigue
A signal only reaches the feed after it passes an **alert gate**. It must be urgent, actionable, visible to
you, not already covered by another signal, and not a type your team has shown to be noise. Signals whose
action would have a high blast radius (**B3** or **B4**) always go to a person.
* **De-duplication.** If Hive detects the same situation again, it updates the existing signal instead of
creating a new one. The row then reads, for example, "seen 2m ago · 47th time · first raised 9d ago".
* **Grouping.** When several related signals fire together (an alert storm), Hive shows one row with a count
instead of many rows.
* **Cooldown.** Hive does not re-notify about the same thing repeatedly, but breaks the cooldown early if the
severity gets worse.
* **Alert fatigue.** If your team dismisses a type of signal at least 80% of the time across at least five
signals in the last 14 days, Hive stops notifying about that type. This lifts on its own as those dismissals
age out. See [Triage signals](/signals/triage-signals) for how each dismissal reason teaches Hive.
A quiet feed does not automatically mean all is well. It can also mean a source is delayed, a permission is
missing, or a metric is still learning its baseline. Check [Coverage](/signals/coverage) before treating
silence as reassurance.
## Evidence and provenance
Each reading names its source connector and field, and can show the baseline it was compared against.
Comparisons are labelled in plain words — "past your threshold", "against your usual", "corroborated",
"correlated" or "cited" — so you can tell a measured breach from a supporting observation. Source links always
point to the original record over a secure link and never contain credentials.
## What happens when you disconnect a tool
When someone disconnects an account, Hive stands down the open signals that account was feeding and removes
their pending approvals. Nothing that was already approved is undone.
## Related
Act on, snooze or dismiss a signal, and what each choice changes.
See what Hive is watching in each tool.
Decide on the actions Hive proposes.
The principles behind every signal Hive raises.