On-call has been the last largely unautomated workflow in engineering. A 2026 Honeycomb survey of 1,200 SREs found 38% are actively shrinking their rotations because of AI-assisted triage — and most engineering orgs haven't yet thought through what happens to the humans, the shift structure, and the career ladder once that's real.
What AI on-call actually does today (and doesn't)
Production AI reliably handles initial alert triage across logs, metrics, and traces; runbook execution for known incident patterns; drafting incident communications; and deciding whether an incident needs a human at all. What it still can't do reliably: novel root-cause analysis, causal reasoning across seemingly unrelated systems, or judgment calls about customer impact. That split — mature at routine triage, immature at genuine novelty — is exactly what's driving the restructuring below.
The three-tier response model
- Tier 1 — agent triages. The agent correlates signals, checks recent deploys, matches known patterns, and either resolves the incident or escalates. The human never wakes up.
- Tier 2 — agent escalates with context. If Tier 1 can't resolve it, the human gets a curated brief: what's affected, what the agent already tried, and the top remediation paths ranked by confidence — not just a bare page.
- Tier 3 — human decides. The human isn't discovering the problem anymore. They're choosing between agent-suggested options or diagnosing something genuinely novel the agent couldn't pattern-match. On-call becomes decision work, not detection work.
Teams running mature pilots of this model report Tier-1 alert volume dropping 60–80% — the incidents that do reach a human are the ones that genuinely needed one.
The redesign most orgs are skipping
If 70% of the noise disappears, rotation size can shrink — that's the honest math, and it's what's showing up in the Honeycomb data. But smaller doesn't mean easier: the incidents that survive the AI triage layer are harder on average, because the easy ones never reach the human. That argues for fewer people with a higher senior-to-junior ratio, longer engagement per incident, more architecture-level context per rotation, and a different fatigue pattern — fewer pages, but each one deeper. Most orgs are shrinking the rotation without redesigning the shift itself, and that's the mistake: a smaller rotation with the old shift structure burns people out faster, not slower.
Why this matters for platform teams
If you're piloting AI-assisted triage, the organizational redesign deserves the same planning cycle as the technical rollout — decide the new rotation shape and seniority mix before headcount shrinks, not after the first burnout cycle proves the old shift structure doesn't fit the new incident mix.
The full issue on LinkedIn goes deeper into what "agent-executable" runbooks look like, in contrast to the human-readable ones most teams still have.