AI incident assist

AI incident assist turns an incident's monitoring data into writing you would otherwise do by hand: a summary for whoever joins the incident, a customer-facing status page update, and a postmortem draft. Everything it writes is a draft. Nothing is published to your status page or saved as your postmortem until a person on your team approves it.

What's included#

AI assist lives on each incident's page (Incidents in the sidebar, then click an incident).

The facts in Incident insights (where it failed, error types, response times, deploys and related incidents) are computed from your data without AI and appear on every plan. Deploy matches need deployment correlation, which depends on your plan. See Plans & limits for what your plan includes.

Monthly AI allowance#

Each plan with AI features includes a number of AI generations per month. The number depends on your plan: see Plans & limits. The panels show your usage as AI drafts: used / limit this month.

• Every successful summary (including automatic ones), status update draft, translation and postmortem draft counts as one generation.
• If nothing has changed since an earlier generation of the same kind in the last 30 days, SutramX reuses the stored result and it doesn't count. You see "Nothing changed since the last summary, so it was reused." for summaries.
• A request that fails on the AI side doesn't count.
• The allowance resets on the 1st of each month (UTC). Unused generations don't carry over.
• AI features are included in the plan price; there is no separate charge.

When you've used your allowance, you get "You have used all N AI drafts included in your plan this month. The allowance resets on the 1st (UTC)." You can still write everything yourself. Generation requests are also limited to 30 per 10 minutes per workspace, to protect against accidental bursts.

AI summary#

The summary helps someone who just joined an incident understand it in about a minute.

1. Open the incident.
2. In Incident insights, click Summarize.

You often won't need to click: once an incident has been open for 5 minutes, SutramX writes the summary automatically so it's ready when someone opens the incident. It then says "Written by AI automatically". This applies to incidents that aren't suppressed, uses your monthly allowance, never uses the last 20% of it (that part is kept for your team's own requests), and is limited to 10 automatic summaries per workspace per day.

The summary has a headline and a short overview, followed by Impact, Regions, Errors, Deploys (when there are deploy matches), Likely cause with a confidence of low, medium or high, and Next steps. When the evidence doesn't point anywhere, the likely cause says so instead of guessing.

The summary is stored on the incident, so your teammates see the same one. If new checks or events arrive after it was written, it shows "New data has arrived since; refresh for an updated summary." Click Refresh to generate a new one.

AI status page updates#

Draft a customer-ready update for your status page, review it, then publish it.

1. Open the incident and go to Status page updates.
2. Optionally pick a Stage: Investigating, Identified, Monitoring or Resolved. Leave it on Pick for me to let the AI choose from the incident's state.
3. Optionally fill in What customers should know (up to 1,000 characters), for example "We found the cause and are rolling back a release". This is the only way to give the AI a cause to mention.
4. Click Draft an update.
5. Review and edit the text of the draft (up to 5,000 characters). Click Save draft to keep edits without publishing.
6. Click Publish and confirm.

Published updates become public notes on the incident: they appear under the incident on every status page that shows this monitor, and in the incident's Notes. A published draft is shown with its publish time; it can't be edited or published again. Click Discard to drop a draft you don't need.

Drafts are written for your customers, not your engineers. The AI is told to:

• Keep it to one to three short sentences that say which service is affected and what customers may notice.
• Leave out internal details such as hostnames, IP addresses, error codes, region codes, deploy names, team names and people.
• Not speculate about the cause unless your guidance or an earlier public update states it.
• Not promise a resolution time.

The stage guides the wording of the draft; your status page shows the text you publish.

If the monitor isn't on any status page yet, the panel says so. You can still draft and publish; the update appears once you add the monitor to a page. To post an update without AI, add a public note instead: see Incidents.

Translations#

If a status page that shows this monitor offers extra languages (see Status pages overview), the update is drafted in the page language and translated into each extra language at the same time. Each draft then has a text box per language.

• Edit any translation before publishing. Leave a box empty and visitors in that language see the main text.
• If you edit the main text, click Translate to translate the current text again. This counts as one generation.
• When you publish, the translations are published with the update.

For public notes you wrote yourself, open Translations of public updates under Status page updates, expand a note, fill in each language and click Save translations. These manual translations don't use AI.

Postmortem drafts#

On Pro, resolved incidents get a Postmortem draft panel next to the postmortem editor.

1. Open a resolved incident.
2. In Postmortem draft, click Draft postmortem.
3. Edit the draft. It has a Title, Summary, Impact, Timeline (UTC), Root-cause hypotheses (each with evidence and a confidence of High, Medium or Low), Contributing factors, Action items (each with a priority and an owner role) and Lessons learned. Add or remove entries with the Add links and bin icons.
4. Click Save draft to keep your edits.
5. Click Use as postmortem and confirm. The draft is converted to Markdown and replaces the incident's current postmortem text.

After that, the postmortem is regular text you can keep editing in the Postmortem section. The draft is blameless by design: it describes systems and decisions, and action items name a team role (for example "on-call engineer"), never a person.

To start over, click Redraft. If you edited the draft, you're asked to confirm first, because a new draft replaces your edits.

What data is sent to the AI#

AI drafts are generated by OpenAI through the OpenAI API. Status page updates, translations and postmortem drafts are generated only when someone on your team clicks a button. Summaries are also written automatically for incidents that stay open (see AI summary).

The AI receives a fact sheet built from the incident, after redaction:

Which notes are included depends on what is being written:

• Summaries and postmortem drafts include internal and public notes, because they are for your team.
• Status page updates include only public notes, so internal notes can't leak into a customer update.
• Translations send only the text of the update being translated.

Before anything is sent, free text (error messages, notes, your guidance) is redacted: query strings and fragments are removed from URLs, email addresses become [email], phone numbers become [phone], long token-like strings become [redacted], and bearer tokens, JWTs and secret-looking values are stripped. For our list of service providers, see the privacy policy.

Review before publishing#

AI assist never publishes anything on its own:

• Summaries, including automatic ones, are shown only in your dashboard.
• Status page updates stay drafts until someone clicks Publish.
• Postmortem drafts stay separate until someone clicks Use as postmortem, and the postmortem is never public.

Read every draft before you publish it. The AI only knows what your monitoring data shows, so check names, times and anything it says about cause or impact.

Troubleshooting#