Draft Incident Comms and On-Call Handoffs with a Human Send Gate

Tier: Foundation · Hands-on time: ~30 minutes

Both are summarization over artifacts you already have — timeline, impact data, pager exports — which is exactly where the evidence is strongest: Google's SRE org reports exec summaries prepared in 52% less time with quality on par with human-written. The invariant this playbook enforces: AI drafts, a human sends — always. A wrong status-page update is itself an incident.

When to use

• •
  During an incident, when the comms lead owes three different audiences updates on different cadences and the responders need to keep responding.
• •
  End of an on-call shift (especially follow-the-sun handoffs between India and US/EU rotations) or the weekly ops review.
• •
  Post-incident customer notifications where tone matters and engineering time is scarce.

Skip AI when the update is one sentence ("Investigating elevated errors; next update 30 min") — templates beat models for that.

Prerequisites

• •
  Source artifacts as text: incident channel export or timeline (incident.io/PagerDuty timelines export cleanly), impact numbers (error rates, affected regions/customers if known), and for handoffs the pager log + open-incident list for the shift.
• •
  Any general-purpose LLM; no incident product required.
• •
  Your org's comms rules: what may be promised publicly, legal/brand phrases to avoid, severity naming conventions.

The workflow

1. 1.

   Assemble the facts block. Paste the timeline and impact data under === FACTS ===, and your comms rules under === RULES ===. The model must draft only from FACTS — everything else in the draft is a defect.

2. 2.

   Request all audiences in one shot, with citations and gaps.

   Copy

   Draft three communications from FACTS only, following RULES:
   1. STATUS PAGE (external, ~80 words): impact in customer terms, what we're
      doing, next-update time. No root-cause speculation, no blame, no promises
      of fix times unless present in FACTS.
   2. EXEC BRIEF (internal, ~150 words): severity, customer/revenue impact,
      current hypothesis (labeled as hypothesis), owner, next milestone.
   3. CUSTOMER EMAIL (affected accounts, ~120 words): plain language, what
      happened, what we did, what we'll do.
   After each draft, list which FACTS lines each sentence relies on, and list
   any sentence with NO supporting fact as GAP so a human can fill or delete it.
   

3. 3.

   Audit before send. Check every number, timestamp, and scope claim ("all users," "EU region only") against FACTS verbatim — scope inflation is the classic AI comms failure. Resolve every GAP by deleting or by adding a verified fact. Then the comms lead sends from their own hands: paste into Statuspage, send the email, post the brief.

4. 4.

   For handoffs, swap the output spec. Same FACTS/RULES structure over the shift's pager log:

   Copy

   Produce an on-call handoff: (1) open incidents with current state + next
   action + owner, (2) pages that fired and their dispositions, (3) changes/
   deploys during shift, (4) watch items with the metric link, (5) toil worth
   automating. Every item carries its source link from FACTS. Unknowns say
   UNKNOWN — do not smooth them over.
   

   The outgoing engineer verifies each open-incident state is current (states go stale within the hour), then posts it as their handoff and stays available for one question.

5. 5.

   Reuse the structure, not the text. Save the prompts as team snippets; each new incident gets fresh FACTS. Never let a model "update the previous message" without the updated FACTS — that is how status pages drift from reality.

Tool variants

• Plain LLM chat: paste-and-draft as above; works for every team, free tier sufficient.
• incident.io / PagerDuty Scribe Agent / Statuspage AI assist: generate updates directly from their own timeline data — lower assembly cost, same audit obligation: verify numbers and scope against the timeline before publishing, because the integration does not make the model honest.
• Scripted weekly review: a cron job that pulls PagerDuty/Alertmanager exports, runs the handoff prompt via API, and posts a draft to the team channel Friday morning tagged DRAFT — verify before circulating.

Verification

• •
  Citation pass: every sentence in the sent version maps to a FACTS line you can point at; every GAP was resolved by a human, not by re-prompting until the gap disappeared.
• •
  Numbers, timestamps, and scope words ("all," "some," "only") checked character-for-character against source artifacts — these three categories cover nearly all comms hallucinations.
• •
  Tone check against RULES: no speculation on the status page, hypothesis labeled as hypothesis in the exec brief, no unapproved commitments anywhere.
• •
  For handoffs: the incoming engineer confirms (verbally or emoji-ack) that open-incident states matched reality at handoff time — this closes the loop and catches stale drafts.

Pitfalls

• Speculative root cause leaking into public comms. Models love causal narrative. The status-page rule "no root-cause speculation" must be in RULES every time, and you must still check.
• Scope inflation/deflation. "Some customers in one region" becomes "customers may be experiencing issues" — legally and reputationally different. Audit scope words hardest.
• Auto-sent drafts. Wiring the API script to post directly to a public channel or status page deletes the entire safety design. Draft surfaces must be internal.
• Handoff optimism. Models summarize an ugly shift as "mostly quiet with a few pages." The UNKNOWN rule and per-item source links are the antidote; an oncall handoff that hides mess transfers risk to a colleague asleep in another timezone.

Restricted-environment variant

• •
  Client-facing comms in services companies are usually contractually owned by the client's comms process — your AI draft is an input to THEIR approval chain; label it as a draft and never send directly to end customers.
• •
  Redact customer names/IDs from FACTS before using a hosted model; for pager exports (which embed hostnames and account IDs), run the drafting on a local model via Ollama — handoff quality at 8B scale is acceptable because the structure does the work.
• •
  Keep drafts and FACTS inside the client's Slack/tenant when the MSA requires it; your personal AI account is not an approved processor.

Success criteria

• •
  Three-audience comms set drafted, audited, and sent in under 10 minutes during a drill, with zero unsupported claims in the sent versions.
• •
  A shift handoff produced in 5 minutes that the incoming engineer rates as complete — measured by them raising zero "wait, what about X?" questions in the first hour.
• •
  Your team's saved prompts include RULES blocks reviewed by whoever owns comms/legal sign-off.
• •
  One month in: drafting time measurably down (aim for the ~50% the evidence supports) with no comms retraction attributable to an AI draft.