The pager goes off at 3 a.m. The engineer on call opens a laptop and spends the first twenty minutes doing the same things as last time: checking dashboards, reading recent deploys, searching logs and posting an update to say they are looking into it.
Most of that work is necessary, but very little of it needs a person who has just been woken up. That is the gap an AI teammate on the rotation is meant to fill. It takes the first pass, gathers the evidence and hands the engineer a clear starting point, so human attention goes to decisions rather than searching.
This guide covers what that looks like in practice: which tasks it takes first, how the handoff works, what changes in the rotation and how to measure whether it is helping.
Key takeaways
It takes the first pass, not the pager. An AI teammate triages alerts, gathers context and drafts a diagnosis; an engineer still owns the incident and any risky change.
Start with routine, low-risk work. Alert triage, evidence gathering, status updates and known runbook steps are good first tasks.
The handoff is the product. A useful teammate hands over evidence, hypotheses and a proposed next step, not just a summary.
Measure engineer effort, not just speed. Track interruptions, time to first useful diagnosis and how often its findings hold up, alongside resolution time.
What an AI SRE Teammate Is
An AI SRE teammate is an AI agent that joins your on-call workflow and does part of the reliability work a human SRE would do: reading alerts, investigating across your tools, drafting a diagnosis and preparing fixes for review. It works where the team already works, such as the incident channel and the ticket queue, rather than in a separate console nobody checks at night.
The word that matters is teammate. A dashboard waits for someone to look at it. Traditional SRE automation tools run the same predefined steps every time a condition is met. A teammate picks up work, follows the evidence, reports back and knows when to hand over to a person. It operates inside permissions your team sets, and its work is visible and reviewable like anyone else's.
It does not replace the engineer on call. Accountability for production, customer impact and risky changes stays with people. What changes is how much of the groundwork is done before a person has to step in. For a closer look at the investigation itself, see how an AI SRE agent investigates, fixes and verifies.
Which On-Call Tasks the Teammate Takes First
The best first tasks are frequent, well understood and safe to get wrong in a draft. They usually fall into five groups.
- Alert triage. Check whether an alert is new, a duplicate or a known flapping check, and group related alerts into one incident instead of five pages.
- Evidence gathering. Pull the relevant metrics, logs, traces, recent deploys and configuration changes into one place, with links to the source.
- First diagnosis. Compare the symptom against recent changes and dependencies, list likely causes and say what would confirm or rule each one out.
- Status updates. Draft the incident channel update and the timeline, so the engineer does not have to stop investigating to write prose.
- Known runbook steps. Run read-only checks from an existing runbook, and prepare the next step for approval where it changes something.
Notice what is missing from that list: unreviewed production changes, access changes and anything destructive. Those come later, if at all, and only for a task class the team has watched work reliably.
How the Handoff Between the Teammate and the Engineer Works
The handoff decides whether the teammate saves time or adds another thing to read. A good one has three stages.
1. The teammate picks up the alert
When an alert fires, the teammate starts the first pass straight away. It posts in the incident channel that it is investigating, so nobody duplicates the work, and it begins pulling evidence from the systems it can read.
2. It hands over a briefing, not a guess
When it reaches a conclusion, or a limit, it hands over to the engineer on call. The briefing should include:
- What is happening: the symptom, the affected service and the likely customer impact.
- What it checked: each source, with links, so the engineer can verify rather than repeat the work.
- What it thinks: ranked hypotheses, with the evidence for and against each one.
- What it proposes: a next step, its risk and how to undo it.
- What it could not do: missing access, missing data or questions only a person can answer.
3. The engineer decides
The engineer reviews the briefing, asks follow-up questions and decides what happens next. If the proposal is a change, it goes through the same review and approval as any other change. At dave.io, for example, the agent handles first-touch triage and drafts a root-cause analysis and course of action, while a human Solutions Engineer owns escalation and any change that carries real risk.
The same handoff works in reverse. The engineer can hand routine follow-up back to the teammate: gather more data, draft the post-incident timeline, or open a ticket for the permanent fix. Our guide to how a DevOps AI agent fits into an existing CI/CD pipeline covers how those follow-up changes move through review.
What Changes in the On-Call Rotation
The rotation itself usually stays the same. Someone is still on call, escalation paths still apply and people still own services. What changes is what the shift feels like.
Fewer cold starts
The engineer rarely starts from a bare alert. They start from a briefing that already says what changed and where to look. That matters most at night, when context-switching is hardest.
Less noise reaches people
Duplicate alerts, known flapping checks and issues that resolve themselves can be grouped, annotated or closed with a note, according to rules the team agrees. That only works if the rules are explicit and reviewed. The teammate should never quietly suppress alerts on its own judgment.
Different work for the people on call
With the first pass handled, engineers spend more of their shift on judgment: deciding between fixes, talking to affected teams and following up on root causes. Some teams find that this makes on-call easier to share with people who are newer to a system, because the briefing carries context they would otherwise have to ask for.
New things to own
Someone needs to own the teammate's configuration: what it can access, which runbooks it follows, and how its rules change over time. Treat that like any other piece of on-call tooling, with an owner and a review cadence.
How to Measure Whether It Is Helping
Pick a few measures before you start, record a baseline from recent incidents and compare like with like. Useful ones include:
- Pages that reach a person. Are fewer alerts waking someone up, without real incidents being missed?
- Time to a useful diagnosis. How long from the alert to a hypothesis the engineer agrees is worth acting on?
- Mean time to resolve. Track it, but read it alongside incident severity and mix. A quiet month can make any tool look good.
- Accuracy of the briefing. How often did the teammate's top hypothesis turn out to be right, and how often did it send someone the wrong way?
- Engineer effort. Ask the people on call. Did the briefings save them work, or did they have to re-check everything?
Be honest about the cost side too. Count the time spent reviewing the teammate's output and correcting it. If reviewing takes longer than doing the work, narrow its scope before widening it.
The goal is a rotation that is easier to carry, not one with a new system to babysit. If you want to see how Dave takes the first pass on your alerts while your team keeps control, book a demo with Dave.
FAQs
Most are built to work in chat, because that is where incidents are coordinated. Dave works in Slack and your ticketing tool. For any product, check which chat tools it supports, whether it can post in private incident channels and how its messages are logged.
It can take the first pass at any hour, which is where it helps most. It should not be the only responder. Keep a human on call for escalation, and agree in advance which situations must always page a person, whatever the teammate concludes.
Mostly from what you connect: repositories, infrastructure code, monitoring, runbooks and past incident notes. Good runbooks and service ownership data make a big difference. Expect early briefings to need correction, and feed those corrections back into runbooks rather than relying on the tool to remember.
Automation tools run predefined steps when a condition is met, such as restarting a pod or scaling a group. A teammate decides which steps are relevant, investigates across systems and explains its reasoning. Many teams use both: automation for known fixes, the teammate for everything around them.
Not much, but agree on a few habits: how to ask it for more evidence, how to reject a proposal, and when to take over. A short walkthrough during a low-severity incident usually works better than documentation. Revisit the habits after the first few real incidents.
Meet Dave
Hand your infrastructure backlog to Dave.
An AI platform engineer, backed by humans. Your team approves every change.
book a demo