> ## Documentation Index
> Fetch the complete documentation index at: https://docs.use.observer/llms.txt
> Use this file to discover all available pages before exploring further.

# AI assist

> Optional drafts of incident updates and postmortems from measured evidence, and translation of published updates. What it sends, what it never does, and how to turn it off.

AI assist drafts text for you to review. It never publishes anything on
its own, and it is optional: you can turn it off for your whole
organization at any time.

AI assist is included on paid plans. It is not available on the Free
plan.

## What it does

### Incident update drafts

**Draft** in the incident editor drafts the next customer update: a
title, severity, status, affected services and a short message. It works
from the incident, its earlier updates and its measured
[evidence](/docs/guides/incident-evidence). A message type or severity
you already chose is kept.

Drafts also work for messages on an existing maintenance window, from the
message dialog.

For an incident that Observer opened automatically, the first draft
starts by itself the first time someone opens the incident, and is kept
for later visits. Nothing is posted.

### Postmortem drafts

**Draft** in an incident's **Postmortem** section fills in the facts from
the incident's own data: impact, timeline, detection, resolution and
every number. These parts are computed in code, not by AI. AI fills only
the summary, root cause, contributing factors and action items, and a
person reviews and saves the postmortem. Postmortem drafts end with the
incident's evidence appendix.

### Translation of published updates

On a status page where you turn on **Translate incident updates**
(page editor, **Identity**, under the page language), messages you
publish in another language are machine translated into the page's
language. This covers incident and maintenance messages.

* Translation happens after you publish. The original is never changed,
  and visitors can always show it.
* Translated messages are marked as machine translated on the page.
* A translation is rejected if it drops or changes any number or URL from
  the original.
* Subscriber notifications wait a short time for the translation, then
  send the original if it is not ready.

Translation is off by default on every page. See
[Page language](/docs/guides/page-language).

## What it never does

* **It never publishes.** Every draft is a suggestion in the editor. A
  person reviews it, edits it and posts it.
* **It never guesses the root cause.** A draft cannot state a cause
  unless an operator update already stated it. A postmortem root cause is
  kept only when it quotes words from your own updates; otherwise the
  field is left for you to write.

Every draft is checked against the measured evidence before you see it.
A draft is refused, and you get no draft, when it:

* makes a causal claim that no operator update made;
* contains an internal identifier, such as an id, a key or a redaction
  marker;
* states a number (a time, duration, count or percentage) that is not in
  the incident or its evidence;
* says the issue has recovered while the evidence shows it ongoing,
  unless you are writing a Resolved update;
* says customers are still affected after a measured recovery.

When a draft is refused, write the update as usual.

## What it sends

AI assist uses Anthropic as its AI provider. Anthropic is listed on
Observer's [subprocessors page](https://use.observer/policies/subprocessors).

| | |
| - | - |
| **Sent to Anthropic** | Incident titles and severity, service, page and check names, the text of incident messages, metric status changes and SLO figures for the incident. Translation sends the published message text. |
| **Removed before sending drafts** | Internal identifiers, probe configuration, metric values, and anything shaped like an email address, IP address, key or token. |
| **Never sent** | Subscriber or end-user contact data. |
| **Anthropic's handling** | Anthropic does not train models on this data, and deletes API inputs and outputs within 30 days unless it must keep them longer to enforce its usage policy or comply with the law. |

Translation runs only on status pages where you turn it on.

## Turn it off

Open **Settings**, then **Organization**, and turn off **Use AI assist in
this organization**. Any member of the organization can turn it off.

Once it is off, only an organization admin can turn it back on. For
other members the switch is locked with "Turned off for this
organization. Only an admin can turn it back on." This way one member
cannot quietly start sending incident text again after someone else
turned AI assist off.

When AI assist is off:

* the **Draft** buttons are hidden and no drafts are made;
* no new translations are made, and existing translations are hidden on
  your pages, in feeds and in notifications, which show the original
  text;
* nothing is sent to Anthropic.

Both changes are recorded in the
[audit log](/docs/reference/audit-log-events), with who made them:
`org.ai_assist_disabled` when it is turned off and
`org.ai_assist_enabled` when an admin turns it back on.

## Monthly allowance

Each paid plan includes a monthly AI assist allowance. **Settings** shows
how much of it is used. When it is used up, drafts stop with "Monthly AI
allowance used" and new translations wait until the allowance resets on
the first day of the next month (UTC). Everything else in Observer keeps
working.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.