For this decision, is the human deciding, checking, or absent — and is that right?
Baseline: 2Enhanced: 13 controls
Who it applies to
Enter for every AI system. Some of its controls apply only where it acts on its own (B3) or it decides about people (B2). Go as deep as your system letter, and implement in your mode.
Why it exists
Choose, for each use case, exactly where the human sits relative to the AI — and drive that choice from the reversibility classification you already made in Governance. There are three positions, and the difference between them is not cosmetic. In one, the human decides and the AI only assists — the AI proposes, the human disposes, and nothing takes effect without the human. In another, the AI acts and the human watches, able to intervene — suited to high volume and lower stakes, where stopping every action would defeat the purpose but a human must be able to catch and halt what goes wrong. In the third, the AI acts alone — appropriate only where the action is fully reversible. Assign the right position to each use, deliberately, and record it.
What to actually do
Assign one of the three positions to every acting use case, from its reversibility. The irreversible actions get a human deciding; the reversible-at-a-cost actions get a human watching and able to stop; only the fully reversible get autonomy. This is the same rule you set in Governance, now turned into an operating decision about where a specific person stands in a specific workflow.
Do not let “human watching” quietly become “human absent”. The position where a human is meant to watch and intervene is the one that decays fastest: over time, with nothing going visibly wrong, the watching stops. Guard against it deliberately — with the interface and competence measures below — because a nominal watcher is autonomy wearing a costume, and it fails exactly when it is finally needed.
Put a person between AI-generated content and the outside world. Whatever leaves the organisation — a published article, a letter or answer to a customer, a report to a client, a public post — is checked by a named person before it goes: for invented facts and sources, for statements the organisation would not stand behind, and for fitness for its audience. This holds whether or not the use decides about people or acts on its own; content does harm by being believed.
Controls
Baseline2 controls
Control 192
Assign an oversight position — human decides, human watches, or autonomous — to every acting use case.
Driven by the reversibility classification, recorded per use case.
How to implement it · Minimum content
Use case and the action the AI takes or proposes, including actions on people
Reversibility, from the existing classification: fully reversible, reversible at a cost, or irreversible
Position: human decides (irreversible), human watches and can stop (reversible at a cost), autonomous (fully reversible only)
Named person or role holding that position
Date assigned and who approved it
How to start
List every use where the AI acts on its own or decides about people, take its reversibility class, and write the position next to it.
Evidence
The oversight position recorded per use case, with its reversibility class.
Mode: light / standard / full
light
A column in the AI inventory spreadsheet, approved by the owner-manager.
standard
Recorded in the inventory per use case and approved by each system owner.
full
Checked at deployment approval, with the position configured in the workflow so it cannot be bypassed.
Control 241
Have a named person verify AI-generated content before it is published or delivered outside the organisation.
BaselinePreventiveall rolesper system
What it is
Checked for accuracy, invented facts and sources, and fitness for the audience; who verified is recorded with the release.
How to implement it · Steps
Scope: anything leaving the organisation — published articles, customer letters and answers, client reports, public posts
Check facts, figures, names and quotations against a source; remove any reference that cannot be found
Check the organisation would stand behind every statement, and that it suits the audience
Record who verified and the date with the release, e.g. initials on the document or a field in the send log
How to start
Tell staff that AI-drafted external content needs a named checker, and add a "verified by" line to the release step for each type of content.
Evidence
Release records naming who verified each piece of AI-generated content sent or published.
Mode: light / standard / full
light
The author or a colleague checks the piece and initials it in the document or the email.
standard
Editorial or client-delivery reviewers sign off in the publication or document workflow.
full
The release workflow blocks publication until the verifier field is completed, with periodic sample checks.
Enhanced1 controls
Control 193
Match the position to the stakes, and re-examine it on change.
EnhancedPreventiveall rolesper system
What it is
Reassessed when the use, the model, or the impact changes.
How to implement it · Minimum rules
Stakes check: beyond reversibility, weigh impact on people, money and reputation; move to a more supervised position where stakes are high
Reassess when the use changes: new purpose, new user group, new type of action
Reassess when the model or provider changes, including provider-announced updates
Reassess when the impact changes: higher volume, higher amounts, more people affected
Each reassessment recorded: trigger, decision kept or changed, approver, date
How to start
Add "re-examine the oversight position" as a mandatory item in the change process for this system.
Evidence
Reassessment records showing the trigger, the decision and who approved it.
Mode: light / standard / full
light
The owner-manager reviews the position when told of a change by the person running the tool.
standard
The system owner reassesses on each change ticket, with the decision recorded in the inventory.
full
Change management blocks release until the reassessment is recorded, with periodic sampling by the governance function.
Evidence
The oversight-position assignment per use case; the mapping from reversibility to position; reassessment records; release records naming who verified AI-generated content published or delivered outside the organisation.