Using ChatGPT for SRE analysis and writing

New requests are succeeding again. Earlier failed jobs still need recovery. A summary that says “the service is restored” has lost the part of the incident people still need to handle. In this hypothetical case, ChatGPT can help turn the observations into a clearer update; giving it permission to close the incident would ask it to do a different job.

That is a useful starting point for SRE adoption. Choose work whose output can be checked against known inputs, then measure preparation and review together. Before supplying workplace records, confirm which account, tools and data classes your organization permits; the final section explains that prerequisite. A permitted timeline and selected observations can become a useful draft without assuming the assistant can continuously monitor the service or decide when its recovery is complete.

From source material to a useful draft

Give the assistant material to work from, a task and an audience. For a runbook, that might mean the existing procedure and a request to explain the branching decisions to a new responder. For a change review, it might mean two configuration versions and a request to describe the differences. For a customer update, it means the observations that may be communicated and a request to turn them into plain language. These are different outputs even when they use the same source.

For a hypothetical runbook revision, suppose the old procedure sends every failed export to replay, while the proposed procedure first checks whether the original job completed. A useful comparison names that new decision and explains why replay may now be skipped. A list saying “added a status check” would miss the behavior the responder needs to understand.

The useful work is the transformation. Scattered observations become an account someone can follow; dense instructions become an explanation of the decision they support. Keep the source nearby so the reviewer can check that change in form did not become a change in meaning. Coherent prose cannot supply an observation that was never there.

A useful request is specific enough to make the result reviewable: “Explain which step changed, what behavior that changes and which details the supplied material does not establish.” The next sections apply that approach to a short incident update, then to the broader analysis and writing work an SRE team already does.

Runbooks, timelines and draft updates

Language work is a good place to see a concrete benefit. A long incident timeline may already contain the answer a new responder needs, scattered among updates written an hour apart. ChatGPT can assemble those observations into a draft, compare two runbook versions or collect unanswered questions. The useful result is less reading and clearer connections, provided the draft preserves the distinctions in the source.

For the hypothetical backlog case, supply two observations: “14:10: new requests are succeeding” and “14:12: failed jobs from before recovery have not been replayed.” An incomplete draft says, “The service is restored.” A supported update says, “New requests are succeeding again; recovery of the earlier failed jobs remains open.” Omitting the second observation changes partial recovery into a broader claim the supplied evidence cannot support.

Closing an incident requires a defined recovered state, permission to change the record and a way to verify that state beyond the generated draft. A successful summary trial can justify continued writing assistance while ticket closure remains contingent on explicit service and backlog checks outside the summary.

For low-stakes drafting, a lightweight review may be sufficient. A helpful writing result cannot justify access to production data or permission to execute a recovery action.

Apply the same distinction to code and query generation. Ask for assumptions and use an isolated test with prepared inputs and expected results that a reviewer can inspect. Valid syntax does not establish that a query selects the intended requests, and a command can succeed in the wrong environment. Checking output and target is part of the task, rather than overhead to exclude from the claimed saving.

Generated training scenarios can also help when clearly labeled as simulations. An experienced reviewer should check their expected behavior so the exercise does not turn a plausible-sounding recovery shortcut into something trainees are taught to repeat.

Keep source review in the drafting workflow

Read diagram description

Example of language assistance: supplied incident records become a draft update. A responder checks the draft against those records before it is used; drafting does not require production write access. Diagram labels: Approved source material: Timeline and relevant observations; ChatGPT draft: Organize and explain the supplied record; Responder review: Check omissions, claims and uncertainty; Reviewed update: Use the existing communication process.

A supplied snapshot does not establish ongoing coverage

A snapshot and a monitoring service do different jobs. ChatGPT may explain what a supplied image or export shows at one moment. Continuous anomaly detection needs incoming data, a way to detect change, evaluation and a route that delivers the finding. A good explanation of yesterday’s snapshot does not create any of that machinery.

A suggested cause likewise remains a hypothesis until supported by the incident evidence. Reading a dashboard may provide another observation; changing a deployment adds execution consequences. The production-agent guide develops the permissions, action records and containment needed for write-capable integrations.

Preparation time and review time belong together

Try the same drafting job with and without assistance, using authorized historical material or constructed cases with known requirements. Include the unresolved backlog rather than choosing only tidy incidents. Ask a reviewer to identify what is healthy, what remains unfinished and what the next update needs to say. That checks whether the draft improves understanding, not merely whether it sounds like an incident update.

Record supported claims, omissions, corrections, review time and whether the result was useful. These observations reveal where preparation became easier and where checking became harder. Measuring only generation speed would leave the team unable to tell whether the full drafting task improved.

OpenAI's evaluation documentation covers defining tasks and checking outputs. A manual rubric or evaluation service can support the comparison, provided the cases and criteria remain stable enough to interpret changes. Repeat important cases and record the configuration: model, prompt and retrieval changes can alter behavior after a successful initial trial.

The account receiving the operational records

Confirm the approved product, account, connected tools and data class before adding workplace material. ChatGPT and an application built on an API are different deployments with different controls. The workplace AI guide provides the questions to resolve for the environment the team actually uses.

For the person taking over the backlog incident, the useful update is the one that saves them from discovering the unfinished jobs again. It says new requests work and earlier recovery remains open. If ChatGPT makes that account clearer with less total preparation and review, it has done worthwhile SRE work. Closing the incident still belongs to evidence about the service and those outstanding jobs.