← Trust Pack
For CISO · Risk Officer · GC

AI Incident Tabletop Kit.

Six hypothetical scenarios plus a shared five-step response checklist for facilitating an AI incident tabletop. Run it quarterly with your incident-response team. Scenarios are generic by design — no real company names, no real incidents. The point is to surface where your runbook is thin, not to retell someone else's bad day.

Jump to scenarios Back to Trust Pack

Shared response checklist

Apply this 5-step sequence to every scenario. The depth of each step is what the tabletop tests.

Step 1

Detect

  • Who notices first?
  • How do they escalate?
  • Time-to-detect target?
Step 2

Contain

  • What gets paused, immediately?
  • Who has authority to pause?
  • What's the blast radius?
Step 3

Notify

  • Internal: legal, exec, comms?
  • External: customer, regulator, vendor?
  • What clock starts?
Step 4

Document

  • What evidence to preserve?
  • Where does the log live?
  • Who timestamps decisions?
Step 5

Review

  • Post-mortem owner?
  • What policy gap was exposed?
  • What changes by next quarter?
Scenario A

Sensitive data pasted into a public AI tool

A finance analyst pastes an unreleased earnings draft into a public chatbot to "help summarize the highlights." Three hours later, a teammate sees the same phrasing surface in a peer's brainstorm — flagging that the prompt may have been used in training or cached output.

Initial responder

SOC analyst on call · pages CISO + GC within 15 min

Decision points
  • Was the tool on the approved list?
  • Is the content material non-public information? (CFO + GC adjudicate)
  • SEC/regulatory notification clock — does this start now?
  • Vendor data-deletion request — submit immediately?
  • Who tells the analyst, when, and how — coaching or HR?
Post-incident questions
  • What policy did the analyst think applied? Was training current?
  • Why was there no technical block on this tool?
  • Do other teams have the same gap?
Scenario B

AI-drafted email creates legal exposure

A sales rep sends a customer a contract response generated by an AI assistant. The AI inferred terms that contradict the master agreement. The customer's procurement team forwards the email to their legal counsel and asks for written clarification — implying the inferred terms are binding.

Initial responder

Customer's CSM escalates to GC + sales VP within 1 hour of receiving the legal forward

Decision points
  • Do we retract the email, correct it, or stand behind it?
  • Who responds to the customer's counsel — sales or GC?
  • Are other reps using the same AI assistant with similar risk?
  • Disable the assistant immediately or scope the disable to contract responses?
  • What goes in writing to the customer; what stays verbal?
Post-incident questions
  • Was the AI assistant approved for contract-adjacent work?
  • What controls would have caught this before send?
  • Notification to E&O insurer — required, prudent, or no?
Scenario C

Customer-facing chatbot gives wrong advice

Your support chatbot tells a customer that a regulated product feature works in a way it does not. The customer acts on the advice, suffers a loss, and posts a screenshot publicly. The post is reshared 4,000 times before the support team sees it.

Initial responder

Head of Support + Comms within 30 min of social-listening alert

Decision points
  • Public response — apology, correction, or silence pending review?
  • Pause chatbot entirely, restrict to certain topics, or keep running with disclaimer?
  • Reach out to the affected customer privately — make whole or refer to dispute process?
  • Is there a regulator that needs to hear from us proactively?
  • Audit log review — how many other customers got similar wrong advice?
Post-incident questions
  • What was the human-review threshold? Was it set correctly?
  • Were known limitations of the bot documented in the AI System Card?
  • Should this category of question route to human by default going forward?
Scenario D

AI vendor outage during critical workflow

A core vendor that hosts the model behind your fraud-detection workflow has a 6-hour outage during your monthly close. Transactions queue up. The vendor's status page acknowledges; ETA-to-restore keeps slipping.

Initial responder

SRE on-call escalates to ops VP within 30 min of detection

Decision points
  • Fall back to manual review, conservative rule-set, or pause new transactions?
  • What's the financial cost of each option per hour?
  • Do we proactively notify customers of slower processing?
  • Is there a contractual SLA trigger — and what does it pay?
  • At what hour-mark do we begin contingency procurement of an alternative vendor?
Post-incident questions
  • Did we have a documented fallback before this happened?
  • Is vendor concentration risk acceptable? Multi-vendor strategy review?
  • Should this workflow be eligible for AI in the first place?
Scenario E

Shadow AI discovery surfaces unapproved tool with PII

A Shadow AI Discovery sweep finds a marketing team has been using an unapproved AI personalization tool for six months. The tool ingested customer-segment data including email + behavioral attributes — categories the team didn't realize were PII under your DPA.

Initial responder

Security analyst who ran discovery · pages CISO + DPO within 1 business day

Decision points
  • Disable the tool today, or grant a transition window?
  • Notification: customers, regulators (GDPR / CPRA clocks), partners under DPA?
  • Data-deletion request to the vendor — how do we verify execution?
  • Was anything generated using this data published or shared externally?
  • Marketing team — coaching, training requirement, formal write-up?
Post-incident questions
  • Why didn't the AI Vendor Intake process get triggered?
  • What detection control would have caught this in month one, not month six?
  • Are other teams running with similar unapproved tools right now?
Scenario F

Hallucinated content published externally

A product-marketing piece on your blog cites a fabricated industry statistic that an AI writing tool invented. A subject-matter expert in your space publicly debunks the stat. The piece has been live for three weeks and is the top organic result for the relevant query.

Initial responder

Head of Content + Comms within 1 hour of public debunking

Decision points
  • Correction, retraction, or rewrite — and how is it labeled?
  • Public acknowledgment, quiet edit, or both?
  • Audit: how many other published pieces used the same tool with the same review depth?
  • Reach out to the expert who flagged it — credit, response, both?
  • Update editorial policy on AI-drafted content review depth?
Post-incident questions
  • What was the human-review standard for AI-assisted content? Was it followed?
  • Should some content categories be off-limits to AI drafting entirely?
  • SEO + brand cost of the correction — how do we measure?

How to run this

Cadence: quarterly is the floor. Pick one scenario per session — 90 minutes total. Don't rush through all six; the value is in the disagreements you surface, not the cards you cover.

Who's in the room: at minimum CISO, GC, comms lead, ops/SRE rep, and one engineering lead familiar with the systems involved. For Scenarios B, C, F also include the relevant business function (sales, support, marketing).

Output: a short written record of where the runbook was thin, what decisions had no obvious owner, and what gets fixed before the next session. The point of a tabletop isn't to demonstrate readiness — it's to find the gaps before something real does.

Scenarios are generic by design — no real company names, no real incidents. They cluster around the failure modes most often discussed in NIST AI RMF Map function workshops and ISO/IEC 42001:2023 Annex A.6 (AI lifecycle) implementation guides. None of this is legal advice or a regulatory-compliance substitute.