Skip to the content.

The executive decision-making discipline. AI incidents require defensible decisions under uncertainty, before facts are complete and before the regulatory or customer disclosure window forces a commitment. PB05 specifies the Executive Decision Packet (AI Edition), the CIA+T impact framing that adds Trust as a peer dimension to Confidentiality, Integrity, and Availability, the 4/24/72-hour planning horizon for next steps, and the Approval Receipt discipline that prevents human approval from degrading into rubber-stamping. Read by executives during incidents; designed by the IC and CISO at peace time.

Part of the AI IR Overlay™ framework. See CONTENT_MAP.md for the full repository map.


Playbook 05: Executive Decision-Making With AI in the Loop

During an AI incident, the executive team faces three converging pressures: the technical situation is unfolding, the disclosure window may already be running, and the stakeholders (board, customers, regulators, employees) are asking for an answer that the technical investigation cannot yet defensibly produce. Executives do not need the technical details that the incident commander is working through; they need a structured, AI-aware decision packet that lets them act confidently within uncertainty. PB05 specifies what that packet contains, what cadence it arrives at, what impact dimensions it covers, and what approval discipline it preserves. The Mental Model’s accountability clause and the framework’s response disciplines all flow through the executive layer; PB05 is how that flow becomes operationally legible to the people who must defend the customer’s posture in front of the board, the regulator, and the customer.

Premise

Other playbooks specify what the response team does technically (PB01, PB03, PB06, PB07, PB08, PB09, PB10, PB12, PB21, PB22) and how the response team communicates (PB17). PB05 specifies how the executive team makes decisions during an AI incident, what information they need, in what structure, at what cadence, and under what accountability framing.

This makes PB05 operationally distinct from the rest of the framework’s playbooks in four ways:

Aspect Technical-response playbook Executive decision-making playbook
Primary audience The IC, the on-call responder, the platform engineer The CEO, CFO, COO, CISO, GC, CCO, CHRO, board chair, audit-committee chair
Time horizon First-hour technical actions; 5-business-day hardening SLA First-hour decision packet, 4-hour update cadence, 24-hour and 72-hour planning horizons
Information density Technical-detail-rich; evidence references; tool-call ledger excerpts Decision-relevant; business-language framing; deliberately low technical density
Failure mode Containment delay, evidence loss, response sequence error Decision under incomplete information that creates legal exposure; over-correction that compounds harm; rubber-stamp approval that defeats safety review

AI incidents produce a distinctive executive-decision shape that traditional-IR executive briefings do not address:

The biggest operational problem with AI-incident executive decision-making is not the absence of information; the technical team produces a great deal of information during the response. The problem is the density mismatch: the executive needs structured decision-support, not technical evidence narration. A 20-bullet update with retrieval-trace excerpts, configuration-diff references, and tool-call-ledger samples does not produce a defensible decision; a 5-section AI Decision Packet does.

Mental Model clauses engaged: all four, applied to the executive layer. The Acts clause produces the Agent Capability Profile section of the decision packet (what the AI could do); the Retrieves and Remembers clauses produce the Provenance Summary (what the AI relied on); the Changes clause produces the change-event evaluation that PB22 (Drift) operationalizes downstream.

Use this playbook when: an AI incident has been declared and the executive team requires the first decision packet · the Materiality and Disclosure call is convening and the convening discipline requires the executive packet structure · the customer is being contacted by a regulator, an auditor, or the press about an AI incident and the executive team requires the briefing structure to support the response · the customer’s board chair, audit-committee chair, or CEO is being briefed out-of-cycle on AI risk posture · the customer is designing a new high-impact AI agent deployment and the executive approval discipline (Approval Receipt) needs to be documented before deployment · the Playbook 18 (Post-Incident Hardening) retrospective surfaces an executive-layer process failure (decisions made on incomplete information, communication misaligned with the technical track, approval that was not informed) · the Playbook 16 (Training) curriculum is being extended to include executive-layer drills.

First-Hour Actions

The first hour of an AI incident’s executive layer is time-boxed by the framework’s response targets: technical containment per Playbook 04 (Tool Design) inside 10 minutes, evidence export per Minimum Evidence Set inside 60 minutes, first stakeholder update per Playbook 17 (Communication) inside 30 minutes. The executive packet runs alongside.

The 60-minute executive triage

Minute Action Owner
0–10 Convene the executive incident channel. The IC notifies the executive team (CEO, CISO, GC, CFO at minimum; additional executives based on incident class) of the incident declaration. The notification is structured: incident class, agent affected, initial scope, containment activated, materiality status (pending, in progress, declared), next-update time. The channel is the operational coordination surface; the formal decision packet follows. Incident Commander + CISO
10–25 Draft the Executive Decision Packet (AI Edition). Five sections (see Evidence Priorities below): Situation (facts only, status-tagged per the Playbook 17 (Communication) Three-Status Taxonomy); Agent Capability Profile (what the AI could do); Provenance Summary (what the AI relied on); Impact (CIA+T framing); Actions Taken and Next Steps (with owners, 4/24/72-hour horizons). The packet is drafted from the customer’s pre-positioned Template Library and customized to the current incident. IC + CISO + Communications
25–35 Apply the Three-Status Taxonomy to every claim in the packet. Per the Playbook 17 (Communication) discipline: Confirmed, Suspected, or Validating. The executive team is making decisions under uncertainty; the status tags name the uncertainty so the decisions can be made defensibly. Untagged claims are removed or tagged. IC + Legal
35–45 Convene the Materiality and Disclosure call per framework/04-materiality-and-disclosure.md if any of the convening triggers apply (customer data touched, regulated data touched, external recipients affected, financial actions involved, customer-facing trust impact identified). The Materiality call is structured around the Executive Decision Packet; the materiality determination is made with the packet as the documentation artifact. IC + Legal + Privacy
45–55 Issue the first Executive Decision Packet to the executive team (CEO, CFO, COO, CISO, GC, CCO, CHRO as applicable). The packet is delivered through the executive incident channel and is logged in the customer’s decision-log discipline per Playbook 15 (Records, Retention) as a Type B equivalent for the executive track. IC
55–60 Confirm the 4-hour next-update cadence. The next Executive Decision Packet is scheduled. The executive team understands what to expect when. Cadence missed without proactive extension is itself a process finding for the post-incident retrospective. IC + Communications

Discipline: the Executive Decision Packet is not a status report. It is structured decision-support. The packet’s job is to enable the executive team to make defensible decisions about disclosure, customer communication, regulator engagement, and operational continuity. A packet that buries the decisions in technical detail produces deferred decisions; a packet that omits the technical anchors produces undefendable decisions. The 5-section structure is calibrated for both.

Critical rule: every Executive Decision Packet is reviewed by Legal before issuance during an active incident. The discipline is not bureaucratic; the packet’s claims may be referenced in regulatory disclosure, customer communication, or downstream legal proceedings, and the Three-Status Taxonomy tags constrain the customer’s later reframing options. The two-pass discipline (Template-Library structure pre-approved at design time; incident-time customization reviewed at issuance) is the operational compromise between speed and review.

Containment Options

PB05 does not introduce a kill-switch variant because executive decision-making is a discipline, not a containment surface. The framework’s existing Kill-Switch Modes apply unchanged to the operational containment. PB05’s containment-equivalent discipline is the decision-scope containment: the actions that bound executive decision-making to defensible scope while the response proceeds.

Decision-scope containment

Action Use when What changes
Status taxonomy enforcement on the packet Draft packets contain unhedged claims that exceed the available evidence All claims are re-tagged with Confirmed, Suspected, or Validating; unhedged claims either acquire a status tag or are removed; the executive team makes decisions on tagged claims rather than implicit speculation
Materiality determination scoping The materiality call is being convened and the determination scope is unclear The Materiality and Disclosure call’s scope is named explicitly: which incident facts, which evidence types, which stakeholder impact dimensions, which regulatory or contractual disclosure obligations. An undefined scope produces an indeterminate determination
Approval-Receipt enforcement on the action plan The packet’s “Next Steps” section includes high-impact actions that require human approval (per PB04 Tool Tiering T2, PB17 Communication external statement, financial action, customer notification) The action is gated by the Approval Receipt discipline: the approver sees the change preview, the destination and domain, the source citations, and the object count or cap before approving. Rubber-stamp approvals are blocked structurally
Two-trap callout activation The customer is under reputational pressure (press inquiry, customer escalation, social-media coverage) and the executive instinct is to act fast The two reputational traps (over-correcting publicly with incomplete information; re-enabling systems without proper re-qualification) are named explicitly in the packet; the executive’s decision is informed by the trap framing rather than just the current pressure
Decision-log lock A high-impact decision is being made under time pressure and the customer’s post-incident defensibility depends on the rationale being captured The decision log records the decision, the rationale, the evidence relied upon, the alternatives considered and rejected, the approver, and the timestamp. The lock means the decision cannot proceed without the log entry; the log is part of the framework’s evidence chain per Playbook 15 (Records, Retention)
Disclosure-window clock activation The Materiality determination has identified that a regulatory or contractual disclosure window has started The clock is started explicitly with the trigger event, the window duration, and the disclosure target (SEC, state regulator, GDPR supervisory authority, HIPAA-covered-entity counterpart, customer per contract). The clock is reviewed at every Executive Decision Packet issuance

The six actions are complementary. The discipline is to be explicit about which decision-scope containment is active and which is dormant for the current incident state.

Evidence Priorities

PB05’s evidence discipline is the Executive Decision Packet (AI Edition). The packet is the operational specification for executive-layer information density during AI incidents; every section has a defined content scope, an approval path, and an integration point with the framework’s broader evidence machinery.

The Executive Decision Packet (AI Edition)

Section 1: Situation (Facts Only)

The packet’s anchor. Plain-language statement of what is happening, status-tagged per the Three-Status Taxonomy from Playbook 17 (Communication).

Content scope:

Length target: 5 to 10 bullets. The Situation section is the executive’s first reading; density beyond 10 bullets defeats the structured-decision-support purpose.

Section 2: Agent Capability Profile (What the AI Could Do)

The “what was authorized” anchor. Defines the agent’s potential blast radius from the customer’s inventory rather than from inference during the incident.

Content scope:

The section defines the **potential blast radius**, not the actual exposure. The actual exposure is in Section 4 (Impact). The Capability Profile says what the agent could have done; the Impact says what evidence suggests the agent actually did.

Section 3: Provenance Summary (What the AI Relied On)

The “what was its context” anchor. Defines what the agent saw and what it remembered, which determines whether the surfaced behavior is isolated or systemic.

Content scope:

The Provenance Summary tells the executive whether the harm is isolated (one anomalous interaction) or systemic (the agent’s context is contaminated and other interactions are also affected). This is one of the most decision-relevant pieces of information in the packet: an isolated incident produces a different disclosure and customer-communication strategy than a systemic one.

Section 4: Impact (Business-Language CIA+T Framing)

The “what does this mean” anchor. Translates the technical situation into business-impact dimensions the executive team can act on.

The CIA+T framing extends the traditional CIA triad with Trust:

Dimension Question AI-incident applicability
Confidentiality Was any data accessed or exposed? Per Playbook 09 (Output Leakage): leakage can occur through authorized outputs without classic breach signals; the assessment must include the output distribution map
Integrity Were records or actions altered? Per Playbook 22 (Drift) and Playbook 06 (Workflow Injection): the agent may have written incorrect records to systems of record; the affected record count and the corrective discipline are the load-bearing details
Availability Were operations disrupted? The containment activation (Mode M3 or M4 per Kill-Switch Modes) itself produces availability impact; the assessment includes both the incident’s direct availability cost and the response’s cost
Trust Was incorrect information sent externally, and to whom? The dimension this playbook elevates to peer status. Trust impact may be the dominant material-harm dimension even when no classic breach has occurred. The assessment must identify the affected stakeholders, the scope of incorrect information, the visibility of the harm, and the trajectory of trust restoration

The Trust dimension is the framework’s recognition that AI incidents produce material harm even when the classic CIA triad shows minimal impact. A finance copilot that recommended an incorrect payment based on a poisoned invoice has produced an Integrity impact (the agent’s recommendation was incorrect) and a Trust impact (the executive who acted on the recommendation has the customer’s trust at stake), but may have produced zero Confidentiality impact and minimal Availability impact. A traditional CIA framing would understate the actual material harm.

Metrology qualification (v0.33.0): Confidentiality, Integrity, and Availability have established quantitative measurement instruments (classification levels, hash/signature checks, SLO numbers respectively). Trust at v0.33.0 is a qualitative impact lens, not a measurable property with peer instrumentation. The Executive Decision Packet captures Trust as the affected-stakeholder count, the identifiability scope, the externally-visible-harm flag, and the trust-restoration trajectory; these are scoping signals, not metric values. A v1.1 candidate is the operational metrology specification that would convert Trust to a measurable peer dimension (see CHANGELOG.md [Unreleased] v1.1 backlog).

Section 5: Actions Taken and Next Steps (With Owners)

The “what are we doing” anchor. Defines the response’s current state and the planned trajectory, with named owners for accountability.

Content scope:

The Approval Receipt discipline

PB05 introduces the Approval Receipt as the discipline that prevents human approval workflows from degrading into rubber-stamping. Every high-impact AI action that requires human approval (per Playbook 04 (Tool Design) Tier-T2 tool tiering, Playbook 17 (Communication) external-statement approval, Materiality and Disclosure regulator-disclosure approval, Playbook 19 (Build vs Buy) procurement approval) must produce an Approval Receipt for the approver before the action proceeds.

Required elements of an Approval Receipt:

Element What the approver sees Why it matters
Change preview (diff) The exact action being proposed, with the before-state and after-state The approver cannot evaluate the change without seeing the change; “approve this email send” is materially different from “approve this email send showing the diff against the agent’s draft”
Destination and domain overview Who or what receives the action’s effect; for emails: the recipient list and domain classification (internal, customer-facing, external, regulated); for writes: the target system, record type, and record count Destination classification is the load-bearing context for the approval; an internal-only action is different from an external-recipient action even when the change content is identical
Source citations and provenance Which documents, retrievals, prompts, or prior memory entries informed the agent’s proposed action The approver can evaluate whether the agent’s information sources are appropriate; an action based on a 6-month-old document or a customer-uploaded document warrants different scrutiny than one based on the customer’s authoritative system of record
Object count or cap How many objects (emails, records, transactions) the action would affect; the cap that triggers escalation An approval for “1 email” is different from an approval for “1,000 emails”; the cap discipline prevents scope creep within an approval

The Approval Receipt’s discipline is structural rather than procedural: the approval workflow technically cannot proceed without the four elements visible to the approver. A workflow that allows approval without the receipt is itself a process finding.

Operational requirement

The first Executive Decision Packet must issue within 60 minutes of incident declaration. Subsequent packets follow the 4-hour cadence until the incident closes or the cadence is explicitly extended. The Approval Receipt discipline applies to every high-impact action in every packet’s “Next Steps” section; an approval workflow that bypasses the Receipt discipline during an active incident is a finding for the post-incident retrospective.

Recovery Sequence

PB05 recovery addresses three scenarios: restoring executive-layer trust after a decision was made on incomplete information, restoring the approval-receipt discipline after a rubber-stamp finding, and restoring brand and credibility after an externally-visible AI incident.

Path 1: Restore executive-layer trust after a decision on incomplete information

A failure mode: an executive decision (disclosure, customer communication, public statement) was made on information later contradicted by evidence. The recovery sequence:

  1. Document the decision and the information state at the time it was made. What information was available, what was tagged Confirmed, Suspected, or Validating, what evidence emerged after the decision. The documentation is honest about what was knowable and what was not.
  2. Apply the Playbook 17 (Communication) Path 1 misstep-recovery discipline for the external-communication dimension. The correction acknowledges the misstep explicitly rather than silently revising.
  3. Brief the executive team on the post-decision evidence. The executive who made the decision is updated directly; the broader executive team is updated through the next Decision Packet’s Situation section with explicit revision tagging.
  4. Update the customer’s executive-decision-discipline artifact if the misstep reveals a structural gap (the Three-Status Taxonomy was not enforced; the packet’s Provenance section was thin; the materiality determination was made on incomplete impact scoping).

Path 2: Restore approval-receipt discipline after a rubber-stamp finding

A failure mode: a post-incident retrospective surfaces that an approval workflow proceeded without the Approval Receipt; the human approval was therefore not informed in the way the framework specifies. The recovery sequence:

  1. Inventory the affected approval workflows. Which workflows allow approval without the Receipt? The inventory is the corrective scope.
  2. Apply the four-element discipline to each workflow. Change preview, destination and domain, source citations, object count or cap. Each workflow either implements the Receipt or is reclassified as a non-approval-gated workflow (which is itself a separate finding requiring evaluation).
  3. Run a Playbook 14 (Testing) drill that exercises the Approval Receipt against a representative high-impact action; the drill validates that the corrective is operational and not just documented.
  4. Update the Playbook 18 (Post-Incident Hardening) backlog with the workflow inventory and the corrective timeline.

Path 3: Restore brand and credibility after an externally-visible AI incident

A failure mode: the AI incident has produced significant external coverage and the customer’s brand is materially affected. The recovery sequence:

  1. Engage the customer’s standard reputation-recovery process per Playbook 17 (Communication) Path 3. PB05’s role is to integrate the executive-layer accountability framing and the framework’s response posture into the broader reputation-recovery effort.
  2. Position the customer’s framework adoption as the credibility anchor. A customer that can demonstrate Approval Receipt discipline, Executive Decision Packet history, Materiality determinations made on documented evidence, and the framework’s Mental Model accountability framing has a defensible posture even after a visible incident. The framework is the customer’s “this is how we are responsible about AI” story.
  3. Run an explicit lessons-learned executive briefing to the board and (where appropriate) to affected customers. The briefing names the specific PB05-discipline hardening items the customer has implemented and the timeline.
  4. Update the customer’s quarterly PB24 board scorecard with the executive-layer findings; the scorecard’s Governance domain absorbs PB05 process findings as part of the customer’s ongoing posture.

Approver for recovery actions: CEO and General Counsel for Path 1 and Path 3; CISO and IC for Path 2. The discipline is the same four-eyes principle from Playbook 15 (Records, Retention): executive-layer recovery decisions warrant senior-leader review.

Post-Incident Hardening

PB05 hardening organizes around four boundaries. Each has an owner, an artifact, and a measurable acceptance criterion. The four boundaries together convert AI-incident executive decision-making from a per-incident scramble into a continuous discipline that the customer’s stakeholders can rely on.

Boundary 1: The Executive Decision Packet codified and rehearsed

Boundary 2: The Approval Receipt discipline operational

Boundary 3: The 4-hour update cadence discipline

Boundary 4: Three executive routine additions

Common Pitfalls

These are the highest-frequency failure modes in AI-incident executive decision-making. Each has been observed often enough to name as a pattern.

Pitfall Why it happens Consequence
Decision packet is a status report, not decision-support The IC defaults to narrating the technical investigation rather than structuring the executive decision The executive team cannot make defensible decisions from the packet; decisions are deferred or made on side-channel information; the customer’s response coherence is compromised
CIA framing without Trust The executive team applies the traditional CIA triad and treats Trust as a marketing concern Material trust impact is undercounted; the customer’s response is calibrated to a smaller harm scope than is warranted; the post-incident review surfaces the gap when stakeholder reactions exceed the customer’s expectations
No Three-Status Taxonomy on packet claims Claims are stated without status tags under time pressure Executive decisions are made on implicit assumptions about the strength of each claim; later evidence contradictions produce credibility costs that explicit tagging would have prevented
Provenance section is thin or absent The technical team has not yet produced retrieval-trace analysis; the IC defaults to leaving Provenance empty The executive team cannot distinguish isolated from systemic incidents; disclosure and communication strategy are calibrated against the wrong incident class
Approval Receipt absent for high-impact action The approval workflow allows approval without the four required elements Human approval becomes rubber-stamping; the post-incident retrospective surfaces that the approver did not have the context to evaluate the action; the customer’s safeguard claim is materially weaker than the customer believed
4-hour cadence slips silently The IC absorbs the cadence as a soft target; subsequent packets arrive at 8, 12, or 18 hours Executive decisions are made on stale information; stakeholder communications drift out of alignment with the technical track; the customer’s response posture appears uncoordinated
Over-correction under reputational pressure The executive team’s instinct under press or social-media pressure is to act fast and broad The over-correction itself becomes a finding; customers are notified prematurely on speculation; regulator filings include unhedged claims that later evidence contradicts
Re-enablement without re-qualification The instinct to restore operations under business pressure outweighs the discipline of validating the response The same incident recurs because the underlying cause was not fully addressed; the customer’s response credibility is materially diminished by the recurrence
No 4/24/72-hour planning horizon The Next Steps section addresses only immediate actions The executive team cannot evaluate the response trajectory; stakeholders cannot be told when to expect substantive updates; the disclosure window may close without the customer’s preparation
Owners not named per action item The Next Steps section lists actions without accountability Actions slip across the response window; the post-incident retrospective surfaces accountability gaps; the customer’s response discipline is compromised
Materiality determination made without the packet The convening call proceeds without the structured information The determination is made on incomplete or inconsistent information; the customer’s documentation of the determination is weaker than the regulatory standard requires
Legal review skipped for time pressure The pre-approval flow is bypassed by the IC during incident operations Packet claims create legal exposure the more deliberate review would have caught; the customer’s posture for regulatory disclosure is compromised
Decision log absent or partial Decisions are made in side-channel communications that are not captured in the formal log The post-incident retrospective cannot reconstruct who decided what when; the regulator’s process review cannot verify the customer’s discipline
No board-briefing readiness The executive team focuses on incident-level decisions without the broader board readout Board members learn about material incidents through external coverage rather than direct communication; the customer’s board-trust posture is compromised; the Playbook 24 (Board-Ready Scorecard) discipline is defeated at the moment it is most needed
No quarterly executive drill The framework’s testing cadence covers technical IR but not executive-layer IR The first AI incident is the first time the customer’s executive discipline is tested under operational pressure; the failure modes from this Pitfalls table all surface at once
Trust impact assessed without the affected-stakeholder count The Trust dimension is named but quantified loosely (e.g., “some customers affected”) The disclosure determination is made on imprecise impact scope; the customer’s notification strategy is calibrated to assumed rather than measured impact

The Question to Carry Forward

If your AI agent caused a customer-visible incident at 14:00 next month, could your IC produce the first Executive Decision Packet within 60 minutes, including all five sections (Situation, Agent Capability Profile, Provenance Summary, Impact with CIA+T framing, Actions Taken and Next Steps with 4/24/72-hour planning horizons)? Could every claim be tagged with Confirmed, Suspected, or Validating? Could the CEO defensibly answer the board’s “what are we doing about this?” question from the packet alone? Could the Trust impact dimension be quantified with affected-stakeholder counts rather than impressions? Could every high-impact action in the Next Steps section satisfy the Approval Receipt discipline (change preview, destination and domain, source citations, object count or cap)?

The honest answer is the gap. If any of those answers is “the IC could not assemble the packet that fast” or “the CEO would need additional briefing first” or “we have approval workflows that proceed without the four required elements”, the Executive Decision Packet template, the Three-Status Taxonomy enforcement, the CIA+T framing, the 4-hour cadence, or the Approval Receipt discipline is the corresponding hardening priority.

AI incidents arrive with operational complexity, regulatory exposure, stakeholder anxiety, and time pressure converging in the same hour. The executive team is making decisions that the customer will defend for months or years afterward, based on information that is partial and changing in real time. The framework’s job is not to slow the executive decision-making to wait for complete facts; it is to make the decisions defensible under the uncertainty that is unavoidable. When the Executive Decision Packet, the CIA+T framing, the 4-hour cadence, and the Approval Receipt are operational, the executive layer becomes a credibility multiplier for the customer’s response posture rather than the gap that surfaces when the first real incident arrives.


Source: AI IR Overlay newsletter, Issue #5, “Executive Decision-Making With AI in the Loop,” by Jacob Ideji. https://www.linkedin.com/in/jacobideji/