ProductAI assist
Incident updates written from evidence.
Observer measures the impact of an incident in code. AI drafts the customer update from that record, never guesses the cause, and a person on your team publishes it.
- Payment success rateUnhealthy2 status changes
- Checkout pageDegraded2 status changes
- acme-prod-eu agent, 2 checks
a41f9c2e07b1checkout-api 2.14.0By timing only: a change near the impact is not shown as its cause.
- Title
- Card payments failing at checkout
- Status
- Investigating
- Severity
- Major
- Services
- Checkout, Payments
Some card payments at checkout have been failing since 14:02 UTC. Customers may see their payment declined or the checkout page load slowly. Our team is investigating and will post an update as we learn more.
The evidence
Measured, not generated.
Every incident gets an evidence appendix, computed by Observer from its own records. No model is involved, and it does not depend on AI being switched on.
Impact window
When it started, and whether it has ended.
Read from the affected checks' status changes, after their dwell time: start, end or still ongoing, duration, worst status and how many separate periods. If no check changed status, the appendix says so and shows the incident's own times.
Affected
Services, checks and customers.
The services marked on the incident, the checks behind their SLOs and each check's status changes. The customers marked affected, and the customers with access to an affected page.
SLO burn
What it cost each objective.
For every SLO that measures those checks: the error budget at the start and at the end of the window, and the burn alerts it raised.
Changes
Listed by timing, never as a cause.
Deploys, commits, releases and Kubernetes events recorded near the start of the impact, each with how far apart they were. A change near the impact is never presented as its cause.
Observers
Who measured it, and from where.
The agents that ran the checks, and how many checks each one ran.
Coverage
A gap reads as a gap.
The share of expected samples in the window that carried a status, so missing data is never mistaken for a healthy reading.
Drafts
AI writes from the record. A person publishes.
Drafts for incident updates and postmortems are written against the evidence appendix and checked against it before you see them.
No guessed cause
It says what the checks show.
Which services are affected, how, since when, and whether they have recovered. Until someone on your team states a cause in an update, a draft uses no causal wording at all.
Checked before you see it
A draft that disagrees with the evidence is dropped.
A number the evidence does not carry, a recovery while impact is ongoing, impact said to continue after every check recovered, an internal identifier: any of these and the draft is discarded instead of shown.
A person publishes
Nothing goes out on its own.
The draft fills the incident editor, marked as an AI draft to review. You edit it, publish it or throw it away. An incident Observer opened on its own gets a suggested draft the first time someone opens it, and stays unpublished until someone publishes it.
Postmortems
The facts are written in code.
Impact, timeline, detection and resolution come from Observer's records, and the draft ends with the evidence appendix. The root cause stays a placeholder unless your own updates stated one.
Languages
In the language of your page.
A draft is written in the language of the status pages the incident appears on. Pages that turn on translation show each published message in their own language.
Ten languages
Every language a status page can use.
When an incident reaches pages in more than one language, the draft is written in English.
- English
- Spanish
- French
- German
- Brazilian Portuguese
- Italian
- Dutch
- Japanese
- Arabic
- Simplified Chinese
Your data
What is sent to Anthropic, and what is not.
AI assist runs on Anthropic's models. Anthropic is listed as a subprocessor, and the list states exactly what it receives.
Sent
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.
Your switch
AI assist runs on paid plans while it is switched on for your organisation. Any member can switch it off in Settings; only an admin can switch it back on.
Usage
Included in your plan.
Every paid plan includes a monthly AI assist allowance in its quota. There are no per-investigation or per-draft fees.
Other AI tools
Your AI SRE finds the cause. Observer proves the impact and tells your customers.
Investigation tools can attach their findings to an incident as internal notes, over the API or the MCP server. Notes stay in the console.
Attach a note
Internal only
Never published.
A note never appears on a status page, in a subscriber email or in a webhook, and adding one never changes the incident. Each note carries the name of the tool that wrote it.
MCP
addIncidentNote
API, Starter+The same operation as an MCP tool, under the key's write:incidents scope.