The 45 toolkit cards

Measure your exposure before automating

Before agentifying anything, measure your degree of dependence, and therefore your exposure: across four dimensions. The greater the dependence, the weaker the sovereignty: govern before you automate.

Map your decisions and their level of autonomy

Map your key decisions and set, for each one, the level of autonomy you delegate to the agent. This is the bridge between the sovereignty diagnostic (ACF-01) and the criticality analysis (ACF-03).

Which decisions can be delegated to an agent?

Objectively determine, for each type of decision, how far to delegate it to an agent: replacing intuition with a reproducible, traceable, and enforceable score, shared across business, technical, and governance teams.

The rules, limits and responsibilities of the agent

Define the agent's "constitution": its identity, the principles that guide it at all times, and the zones that are absolutely off-limits. An agent without a constitution is a risk without a framework.

Frame every deployed agent

Precisely describe an agent in order to set its role, perimeter, data, tools, rules and maturity. The agent card turns the constitution (ACF-04) into a concrete agent that is enforceable and supervisable. 1

Monitor, measure and keep the agent compliant

Place the agent under continuous control: measure its performance, detect drifts, verify its compliance and trace every decision. Supervision transforms a deployed agent (ACF-05) into an agent governed over time.

Emergency shutdown protocol

Be able to immediately shut down an agent or agentic system in the event of drift, incident, loss of control, or critical risk. 1

Document, frame and prepare the deployment of your first agent

Gather in a single document everything that frames an agent before launch: identity, purpose, scope, data, risks, validation. A well-framed agent is a governable agent. 1

Log significant decisions and incidents to ensure traceability and compliance

Ensure the traceability of every significant decision made by an agent · or concerning one. The register documents decisions, validations, incidents, and follow-up actions to guarantee transparency, compliance (AI Act art. 12, DORA), and accountability.

Steer corrective actions and continuous improvement of your agents

Turn an incident, audit or review into a concrete action plan: identify root causes, define corrective and improvement actions, prioritize them (Effort × Impact) and organize follow-up through to closure.

Assess the maturity of your agentic governance and build a 30-day action plan

Measure the maturity of your agentic governance across 8 domains, prioritize actions and execute them over 30 days. Organizational scale 1→5 (CMMI) · distinct from the N0→N3 maturity of an individual agent (ACF-05). 1

Identify, assess and prioritize risks tied to your AI agents to control them more effectively

Identify, assess and prioritize the risks of your AI agents. Each risk is scored Likelihood × Severity (score 1→25), then assigned to one of the 8 ACF domains (same categories as the ACF-11 audit) in order to build a prioritized treatment plan.

Define the scope, limits and responsibilities of every deployed agent

Formalize the mandate of an agent: its mission, its authorization scope, its limits, its level of delegated autonomy, its data, its supervision and its official validation. The mandate is the enforceable reference document binding the agent and those who govern it. (DDAO = Delegated Decision Agent Officer.)

Responsibility by Design · embedding accountability from inception

Ensure that every agentic system embeds, from inception, the five pillars of operational accountability · turning a doctrinal principle into an auditable and enforceable organizational mechanism.

Judging the target against an out-of-reach figure

Write down what the agent optimizes, what it would sacrifice to get there, and which counter-metric watches it without the agent being able to improve it. A perfectly governed agent pursuing a badly designed metric methodically produces decisions nobody wanted.

Testing the arrangement, never the person

Check that the human control that was announced still exists, by measuring what the supervisor sees, how fast it is seen and what is done with it. A control that is signed without being read costs what a control costs and produces none.

Judging whether a mandate still holds

Reopen a mandate on a fixed date and judge its relevance, not its compliance. Checking belongs to the register ; validating asks whether the rules still deserve to be followed. A perfectly obeyed mandate may have become wrong.

Apply the ACF® framework to a real-world scenario, step by step

Put the ACF® framework into practice to analyze a corporate scenario, identify risks, define the appropriate autonomy level, and design the necessary safeguards. This case is completed as a team, card by card.

Facilitate, grade and evaluate agentic governance workshops

Provide instructors with everything they need to run practical workshops based on the ACF® framework, guide participants and evaluate their deliverables. Detailed agendas, teaching cases, grading materials and assessment rubrics.

Train to decide, arbitrate and supervise in real-world conditions

Place participants in a situation of agentic governance facing critical events. Develop the skills of decision-making, arbitration, escalation and supervision within a secure framework, and strengthen alignment with the sovereignty requirements of the ACF® framework.

Who acts, who is accountable, which rule, who stays in the loop

For every act handed to an agent: who performs it, who is accountable for it, which rule bounds it, who stays in the loop. You delegate execution, never accountability.

Placing humans, agents and rules, domain by domain

A one-page overview of how humans, agents and rules share the organisation, and of the domains where autonomy has settled in without governance. It locates; it does not score.

Keeping the ability to judge your agents

The list of skills to keep in-house to stay able to supervise, audit and take back control of your agents. Sovereignty framing, not an HR training plan.

The text that says what is done with AI, before the first agent

Give the organisation a single document covering two things that generic policies rarely separate: the daily use of AI tools, and the moment a system decides in place of a human. It is signed, it is revised, it comes before the first agent.

The tools nobody approved and everybody uses

Flush out the AI tools used outside any framework, sort them, and decide for each whether it enters the policy (ACF-O4) or stops. This card covers tools that assist a person, not agents that decide: those belong to the next card.

The ones the organisation never deployed, already deciding in its name

Find the agents acting in the organisation's name with no mandate: the ones an employee plugged in, the ones a supplier switched on through an update, the ones an open connector lets write. An agent without a mandate binds the organisation all the same.

Who lost a decision, who gained it, with nobody deciding it

Make visible the power that has moved. Every agent deployed takes a decision away from someone and hands another to someone else, usually with that transfer neither discussed nor even noticed.

What was allowed once, and has lasted ever since

Keep the list of departures granted from the rule: who asked, who granted, for how long, and on what condition they close. An exception with no end date becomes the rule, with nobody having had to vote it.

Governing twenty agents is not governing one agent twenty times

Hold a population of agents rather than one agent. List what exists, spot the agents that talk to each other and the shared dependencies, and set the point beyond which the organisation can no longer say what it has deployed.

Whoever bears the decision must be able to challenge it before someone

Name who bears the delegated decisions, how that person learns a system decided, and before which human they can ask for a review. An appeal that leads only back to the same agent is not an appeal.

Setting the agent's six dials

Turn the signed cards into a configured agent. Every dial is copied from a card; nothing is invented at this stage.

Proving every dial before the first customer

Check, in a test environment, that every setting truly holds. An untested setting is a setting you know nothing about.

Pilot → steps → rollback

Bring the agent into production along a reversible ramp, never in one go. You must be able to step back one notch at any moment.

The four rhythms that keep hold of the agent

Supervise an agent without staring at it: a handful of gestures, at a set rhythm. You monitor by having decided in advance what should raise an alert. 1

Who does what, on the ground

Explicitly distribute the running of the agent, otherwise everyone assumes someone else is handling it. R: responsible · A: approves and accountable · C: consulted · I: informed.

Knowing when, and how, to hand over

Settle in advance the situations where the agent stops deciding and calls a human. A cautious agent is worth more than a self-assured one.

The kill switch proven every quarter

Guarantee that the stop works before it is needed. A mechanism never triggered is a mechanism nobody knows works.

Detect → root cause → fix

Ensure no incident passes without a trace, a cause and a dated follow-up. Treat the cause, never the symptom.

Least privilege, reviewed

Periodically check that the agent has only the bare minimum. Every superfluous access is a door an incident cannot take if you never open it. 1

The five questions before buying an agent

Separate a governable agent from a black box. Score each answer out of two points; the total out of ten informs the go / no-go (← ACF-08).

What stops if access stops

Know, before signing, what stops on the day the supplier stops. Most organisations discover at that point that one supplier carries the decision, its record and the means of stopping it.

Deciding without them, then leaving

Measure how long the organisation can decide without its supplier, and what it falls back to. An organisation that cannot decide in degraded mode has not delegated a decision, it has surrendered it.

What reopens the mandate

Tie every signed mandate to the model version the agent was tested on, and reopen the mandate when that version changes. The supplier replaces on its own schedule; the signed text does not move by itself. 1

Evidence you can still produce

The register establishes what was decided, but it must still be producible on the day access stops. Evidence hosted by the party whose behaviour it documents is evidence you can lose. 1

What the framework does not establish

Write down what governance does not establish for this agent, rather than discovering it during an audit. The register proves the act but never its leaning, and that limit is declared. 1