The AI in the Room
Updated
IntermediateWhat you'll learn
~16 min- Use an AI assistant in Review mode for incident drafting tasks, with tools gated to ask, never Autonomous
- Attach the record's ai_drafted marker - draft identifier, evidence referenced, verifier, verification time, accepted changes - to any Timeline or Decisions row an assistant drafted
- State the true Review/Autonomous boundary, and that no verified case of a misleading AI assistant was on record as of 2026-08-28
Before the detail — Decision: let an assistant draft under Review mode with per-tool ask gates, never Autonomous, during a live incident. Outcome: faster first drafts of timelines, summaries, hypotheses, and comms — each one checked by a named person before it enters the record. Artifact: the record’s ai_drafted: marker — draft id, evidence referenced, verifier, verification time, and accepted changes — required by process on every AI-drafted Timeline or Decisions row. validate.py checks completeness and timestamp order once a marker or loose provenance field is present; it cannot detect a fully unmarked AI draft. Status of what follows: reusable guidance.
Prompt first: four drafts, none of them verified yet
Open your workspace agent, switch to Plan mode on your incident logs, chat transcript, and comms template, and start with (if you do not have live exports or files handy, ask it to generate a small Meridian practice sample from the lesson’s scenario first):
From our raw incident logs and chat transcript, draft a source-cited timeline, a next-shift summary, up to three ranked cause hypotheses, and a first status update. Include an unverified ai_drafted provenance stub and flag any missing timestamps.Check that the plan or draft covers:
- Source-cited timeline: one row per event citing the exact log line or message location, marking ambiguous or absent timestamps
MISSINGrather than inferring them. - Three separated companion drafts: (1) a one-paragraph summary for the next shift, (2) up to three ranked cause hypotheses with supporting evidence, and (3) a first customer-facing status update using the approved comms template.
- Unverified
ai_draftedprovenance stub: a proposeddraft_idand exactevidence_referencedlocators ready to paste, while leavingverifier,verified_at, andaccepted_changesblank for human verification.
If the first pass fills in verifier or paraphrases evidence locations, follow up with: "Leave verifier, verified_at, and accepted_changes blank for human review, and cite exact line numbers in evidence_referenced."
Review mode, four drafting tasks
The assistant can read logs, rank hypotheses, and draft a first customer update well inside the calendar time a commander would spend on any one of them alone. Zero Trust 5.2 (a separate training in this series) drew the line this lesson inherits — the assistant triages, a human closes — and Grounded Answers (another training in this series) adds the limit that comes with it: a citation makes a draft defensible, not correct. This lesson adds incident-time provenance: which evidence the draft used, who verified it, when, and what changed. The commander changes incident state, the clock desk records determinations, and the named signer sends required notices.
Reconstructing the timeline itself, in depth, is 5.2’s job, later in this training; drafting the comms update belongs to 4.3’s template. What this lesson owns is narrower and comes first: the mode the assistant runs in while it drafts any of that, and what the record has to say about it afterward.
The record’s provenance: what an AI-drafted row must carry
The record has no separate section for AI output, and it does not need one — any Timeline (section 3) or Decisions-under-uncertainty (section 4) row that an assistant drafted carries a compact marker appended to its own cell, on top of whatever the row already says:
ai_drafted: draft_id=<id>; evidence_referenced=<what it read>;verifier=<human>; verified_at=<UTC instant>;accepted_changes=<what changed, or "none">All five fields (draft identifier, evidence referenced, verifier, verification time, accepted changes) are required whenever the marker appears. The committed NON-PRODUCTION example, INC-2026-041’s own record, carries the worked case at 14:31: an assistant drafted the first reconstruction of the incident timeline from the alert firings and the pool-saturation reading, and the row closes with this excerpt:
ai_drafted: draft_id=ai-2026-041-01; evidence_referenced=sim/replay.excerpt.jsonl#L61-L90, alerts/promtool.test.transcript.txt; verifier=j.whitfield; verified_at=2026-08-18T14:36:00Z; accepted_changes=corrected the deploy timestamp and the first observed 504Five minutes between the draft and j.whitfield’s check, two corrections found and named — the deploy timestamp and the first observed 504 — that gap is the verification the marker exists to prove happened, not just assert.
incident/validate.py checks completeness and timestamp order once a marker or loose provenance field is present: all five fields when the marker appears, a verified_at that is not earlier than the entry’s own timestamp — verification cannot precede the draft — and a row that carries one of these field names as loose text with no ai_drafted: marker around it fails closed, because a dropped tag is not a way to skip verification. The self-test proves the rule holds: 38 mutations, three of them this one, all 38 caught — but the check only runs once a marker or loose provenance field is present; a fully unmarked AI draft is invisible to it, which is why marking is a process rule the assistant is prompted to start, not a safety net the validator supplies on its own.
Writing the marker costs the verifier verification effort per drafted item; skipping it costs a reviewer assessor-hours reconstructing, after the fact, which parts of the record came from a person watching the system and which came from a model reading about it.
One cloud provider’s SRE agent, and what “Review mode” actually means
The categories here are worth more than any one product, but a dated example makes them concrete. One cloud provider’s SRE agent, Microsoft’s Azure SRE Agent (generally available per a Microsoft page dated 2026-08-26), “suggests or, when configured, executes mitigations.” Its two modes are named plainly: “In Review mode, an SRE Agent Administrator approves the write actions that require approval before the agent runs them. In Autonomous mode, the agent applies them without waiting.” Underneath both modes sits a per-tool access policy, “allow, ask, or deny,” but read the defaults closely. As of 2026-11-07, Microsoft’s run-modes page says Review mode shows Approve and Deny “only for Azure infrastructure operations,” and a new incident response plan defaults to Autonomous; its tool-access-policies page says an allow rule runs a tool “immediately, even in Review mode.” Review is a setting you choose and check per plan, not one you inherit.
As of 2026-08-28, Datadog’s fetched Bits Investigation description says it shares findings for engineers to review; treat that as vendor-described behavior, not proof of every capability boundary. Resolve AI says it “autonomously troubleshoots and resolves production issues”; without an independently published method, treat that only as vendor positioning, not a benchmark.
This training teaches Review mode with tools set to ask, verified per plan rather than inherited from a default. Autonomous mode is a real, named setting, not a hypothetical, and it is the door Guarded Automation (Training 09, a separate training in this series) opens, where standing per-action authorization is the whole subject. Whether an assistant may run inside Meridian’s environment at all is a different question still, decided upstream by the authorization boundary in AI Assurance (also a separate training in this series). This is a CROSS-TRAINING HANDOFF, not a federal seam: no clause, no regulator, no determination changes hands here — just a mode setting this training stops short of, and names the training that picks it up. Naming the boundary here is what keeps anyone from discovering mid-incident that “the AI” quietly had standing authority nobody signed off on.
What has not happened yet
State this plainly, because the marketing in the previous section makes it easy to lose: no verified public case exists, as of 2026-08-28, of an AI assistant misleading incident responders during a live response. Every claim circulating about an assistant’s reliability — including the ones quoted two paragraphs up — comes from the party selling the product. That is not a reason to trust the claims; it is a reason to notice that the absence of a documented failure and the presence of a confident one are not the same fact, and only one of them is currently on record.
Teaching the absence honestly, instead of inventing a cautionary story to fill the gap, is the same discipline this lesson has been asking for all along: a hypothesis with no evidence does not become true because it would make a better lesson.
Stop and escalate when anyone on the call proposes moving a tool’s setting from ask to allow, or switching the assistant to Autonomous mode, to move faster mid-incident. That is not the on-call commander’s call to make alone at 03:00 — it is Training 09’s autonomy question and AI Assurance’s authorization boundary, and it routes to whoever owns those decisions outside the incident, not to whoever is fastest to type “yes” on the call.
An assistant ranks three possible causes from cited logs. Without checking the live impact signal, the engineer marks the incident mitigated because the top hypothesis sounds right. What is the problem?
Key takeaway
An assistant in Review mode, with tools gated to ask, can draft a timeline, a shift summary, ranked hypotheses, and a first comms update faster than the room can read the raw logs — and grounding is not truth for any of the four: a citation makes a draft defensible, not correct, and closing actions stay a named person’s to take. The record carries that provenance in the ai_drafted: marker on the row itself — a draft identifier, the evidence referenced, the verifier, the verification time, and what was corrected — rather than in a section of its own, and validate.py fails closed on a marker missing a field or a dropped tag. Autonomous mode is real and named; this training stops at its edge and hands the CROSS-TRAINING HANDOFF to Training 09, while AI Assurance keeps the separate question of whether the assistant may run at all. No verified case of a misleading assistant was on record as of 2026-08-28; say so with the date, and do not invent one. The next module turns from the room during the incident to the clocks that started running the moment it began.
- LEADERSHIP DECISION
- authorize incident-time AI assistance in Review mode only, tools gated to ask and checked per plan for inherited allow rules or an Autonomous default, with every draft the record keeps carrying a named verifier
- PRACTITIONER ACTION
- draft timelines, summaries, hypotheses, and comms with the assistant; verify each against the raw material and attach the ai_drafted marker before it enters the record; never move a tool to allow or the agent to Autonomous mid-incident
- SUCCESS MEASURE
- zero record entries with an AI-drafted origin and no named verifier - every draft either verified and kept, or discarded