AI Governance Tabletop Exercise
A facilitated scenario exercise for leaders: how would your organisation respond if an AI agent acted outside its authority or caused a consequential failure?
Direct answer
Direct answer: AI governance tabletop
A tabletop exercise tests the organisation, not the model. Leaders work through a realistic scenario, such as an agent approving a payment it should have escalated or a customer-facing system giving harmful answers at scale, and discover who notices, who decides, who can stop the system, and what evidence exists. The exercise ends with the response gaps it exposed and a named owner for each one.
The question
“How would our organisation respond if an agent acted outside its authority or produced a consequential failure?”
Who it is for
Boards, risk committees, executive teams, and incident-response owners preparing for AI systems that act on customers, money, or operations.
Governance tabletop: one incident, traced end to end
What you leave with
A scenario pack, a responsibility exercise, the response gaps it reveals, and action owners.
- 01 A scenario pack built around the organisation’s own systems and authority boundaries
- 02 A record of how the room responded: detection, escalation, decision, containment, and communication
- 03 A list of response gaps, such as missing stop paths, unclear owners, and evidence that was not retained
- 04 Named owners and dates for closing each gap
How it runs
Format, method, and preparation
- Duration
- A scoping call, a 2–3 hour facilitated exercise, and a written debrief
- Delivery
- In person in the UK or remote
- Participants
- Six to twenty people: executives, risk, legal or compliance, operations, communications, and the technical owner
Set the scene
Introduce a system the organisation runs or plans to run, its authority, and the controls everyone believes are in place.
Inject the failure
Release the scenario in stages: the first signal, the spreading consequence, and the external pressure, and ask the room what happens next.
Trace responsibility
At each stage, record who would know, who would decide, who could stop the system, and what evidence would exist.
Close the gaps
Agree the gaps the exercise exposed, assign owners, and decide what should be rehearsed again.
Useful to have ready
- →A description of one or two AI systems in use or planned, including what they are allowed to do
- →Existing incident-response and escalation procedures
- →Commitment from the decision-makers who would act in a real incident to attend
Evidence
What this draws on
CloseGate
A Python and MCP policy layer with action tiers, segregation of duties, materiality routing, mandatory approval for irreversible actions, and hash-chained replayable audit. No published independent certification or customer deployment.
Inspect the source ↗Regulus
A runtime-governance architecture covering agent identity and purpose, policy, PII, residency, model-risk tiers, kill switches, human oversight, and evidence export. Controls are mapped to NIST AI RMF and ISO/IEC 42001; mapping is not certification.
Inspect the source ↗Regulated banking AI at Aveni
Architected enterprise banking AI with conduct-risk workflows, evidence generation, human review, escalation, evaluation, versioning, and release controls. Part of the Aveni team in the first FCA Supercharged Sandbox cohort; the FCA lists Aveni as an accepted firm, which is not an FCA endorsement or certification.
Inspect the source ↗Scope and limits
What this is not
- —A tabletop is a learning and decision exercise. Completing one is not certification, an audit, or proof that a system is safe or compliant.
- —Scenarios are illustrative and built from the organisation’s own descriptions; they do not test the live system’s technical controls.
- —Regulatory notification duties and legal positions raised in the exercise are referred to the organisation’s own advisers.
If the need is different
Common questions
Answers before you commission
What is an AI governance tabletop exercise?+
A facilitated discussion in which leaders respond, step by step, to a realistic AI failure scenario. It reveals whether ownership, escalation, stop paths, and evidence work as people assume before a real incident tests them.
What scenarios do you use?+
Scenarios are built around the organisation’s own systems: for example, an agent taking an action beyond its authority, a model change silently degrading decisions, or sensitive data appearing in generated output.
Who should attend an AI incident simulation?+
The people who would actually decide in an incident: the executive owner, risk and compliance, operations, communications, and the technical lead. Board members often attend when the exercise supports board oversight.
Does a tabletop prove our AI governance works?+
No. It shows where the current arrangements would struggle and gives owners a list of gaps to close. It is evidence of preparation, not certification of safety or compliance.
Related