Write Better On-Call Handoff Notes With AI Doing the Drafting

On-call handoff notes are where incidents fall through the cracks - and where AI is making a concrete, measurable difference. Here's what's actually changing.

Cover art for Write Better On-Call Handoff Notes With AI Doing the Drafting

PagerDuty's 2025 State of Digital Operations report found that the average on-call engineer receives roughly 50 alerts per week, but only 2-5% of those require human intervention. That's not the part that breaks teams, though. The part that breaks teams is what happens at shift change - when the exhausted outgoing engineer tries to reconstruct six hours of investigation into a paragraph, and the incoming engineer tries to figure out which threads are still live.

On-call handoff notes are unglamorous, genuinely hard to write well, and consistently under-written. AI is starting to fix that - not by replacing the judgment involved, but by removing the transcription work that gets in the way of it.

Why on-call handoff notes fail at shift boundaries

Shift transitions are operational risk points where knowledge can disappear and incidents can degrade invisibly. Incoming engineers who don't understand current system state waste precious time during new incidents - they rediscover problems the previous shift already identified, rerun diagnostics the previous engineer completed, and repeat investigation steps that yielded no useful information.

The time cost is specific. What could be a five-minute fix becomes a thirty-minute investigation when the incoming engineer doesn't know a recent deployment broke a specific service, or that database replication lag has been gradually increasing for hours.

The structural reason this happens: if a handoff system does not preserve decision context, teams end up re-litigating previous decisions from scratch on every shift boundary. That is tiring, wasteful, and dangerous under fatigue. The incoming engineer should be free to disagree with earlier choices, but they should not have to reverse-engineer them from logs and vibes.

The best handoff process is not the one that looks most complete in a policy document. It is the one that engineers will actually use well at the end of a long shift, during partial incidents, across timezone boundaries, and under the natural cognitive drop that comes with fatigue. That's precisely what makes handoff notes hard to mandate: the people most obligated to write them are the people least able to write them well.

What AI actually contributes to the handoff

The concrete change happening now is that tools are beginning to generate the first draft of the handoff summary from the incident channel itself - not from a form engineers fill out.

PagerDuty's Scribe Agent, which became generally available for Slack users in late 2025, automatically transcribes Zoom calls into the incident chat in real time, then combines that transcript with the chat conversation to draft structured summaries and status updates, so no critical detail is lost during response or in post-incident review.

incident.io takes a similar approach. Their Scribe feature joins incident calls and transcribes in real time, capturing key decisions as they happen. When a new responder joins 30 minutes into an incident, they can run a summary command to get an AI-generated recap of what happened, who's involved, and current status. That last part - surfacing context for someone joining mid-incident - is arguably more valuable than the post-incident review.

After an incident closes, Scribe ensures all the context needed is already captured: timeline events, Slack messages, Zoom transcripts, and responder actions. When you're ready to create an incident retrospective, all that data flows directly into the post-incident review workflow, giving a head start on drafting summaries and identifying action items.

The pattern these tools share: they work from the artifact that already exists (the incident channel, the call transcript, the alert timeline) rather than asking an exhausted engineer to recall what mattered. Rather than engineers spending an hour reconstructing timelines, automation extracts timestamps, actions, and messages. Reports become precise, objective, and consistent. Human effort is reserved for insight rather than transcription.

A teammate like Beagle, living in Slack, can take a similar approach for lighter-weight teams - watching the incident thread and surfacing a draft handoff note at shift end without requiring anyone to open another tool.

The alert noise problem that makes handoffs harder

Even the best handoff note is working against a tide of noise. A 2025 study by Splunk showed 73% of organisations experienced outages linked to ignored alerts. That's not an attention problem. Alert fatigue is the most direct driver of on-call burnout, and it is a systems design problem, not a morale problem

  • an important distinction if you're trying to fix it.

A 2025 industry survey found that 22% of engineering leaders and developers face critical levels of burnout, with another 24% experiencing moderate burnout. On-call responsibilities amplify this problem significantly.

The 2025 SRE Report found that engineers spend a median of 30% of their week on operational work, up from 25% the year before.

When the signal-to-noise ratio is this bad, the handoff note inherits the problem. An outgoing engineer who isn't sure which of their 47 alerts from the last shift actually mattered can't write a tight summary. AI-assisted noise reduction and AI-assisted summarization are, in practice, two parts of the same fix. Pruning noise alerts alone often cuts pager load by 30-40% , which means the handoff has fewer live threads to cover and is more likely to be accurate.

AI can recognize patterns across historic incidents and recommend specific remediation steps. These suggestions resemble having a silent advisor in the room who remembers every similar incident that ever occurred. It shifts response from intuition to informed action.

What still needs a human

The thing AI cannot replace is judgment about what the next shift should actually watch. A generated summary can tell the incoming engineer what happened; it cannot tell them what is about to happen, or flag the thing that was almost an incident but isn't in any alert.

A structured weekly transition meeting - outgoing and incoming engineers both present - should cover active incidents, silenced alerts, and upcoming risky changes. Then the incoming engineer summarizes back before the outgoing engineer signs off. That verbal summary-back step is the part that catches what the written note missed.

The written document is the durable record the incoming engineer can return to at 3 AM. But async-only handoffs lose nuance, tone, and the ability to answer questions - 15 minutes of overlap prevents hours of confusion for the incoming engineer.

The division of labor is becoming clearer: AI handles the reconstruction and transcription; the human handles the risk assessment and the "by the way" context that doesn't live in any alert. The incoming engineer should be free to disagree with earlier choices - they just shouldn't have to reverse-engineer them from logs and vibes.

That's a reasonable place to land. The handoff note was never really a writing problem. It was a cognitive-load problem at the worst possible moment in a shift. Offloading the drafting removes one friction point. The judgment stays where it belongs.

Keep reading