IDD concept: Govern execution

Risk Dials

Risk Dials are governance controls that determine how much autonomy an AI system is granted at each step of a delivery workflow. Each step is set to Red for full human review, Amber for a human spot-check, or Green for autonomous operation with notification, and the organisation moves a dial only when measurement shows it is safe to do so.

The problem Risk Dials solve

The largest barrier to enterprise AI adoption is rarely the technology. It is the fear of losing control. Leaders reasonably ask what happens if an agent makes a critical mistake, how compliance will be maintained and who is accountable when something goes wrong, and without good answers many organisations remain in pilot mode, experimenting widely and scaling very little.

Risk Dials make human-in-the-loop control explicit rather than assumed. Instead of a single, organisation-wide decision about whether to trust AI, there is a separate, visible and reversible decision at every point in the workflow where an agent can act. That turns an all-or-nothing question into a gradual one, and it gives leaders a mechanism for moving from caution to confidence on the strength of evidence.

The three settings

  • Red: high control

    A person must review and approve before the work proceeds. The agent's output is a suggestion, not a decision, and nothing happens without explicit human sign-off. This is where every organisation should start.

  • Amber: medium control

    The agent proceeds and a person reviews a sample of its work, with the ability to override or roll back at any time. The organisation is delegating, not abdicating.

  • Green: low control

    The agent proceeds autonomously and a person is notified and can intervene whenever needed. Green is reached only after trust has been demonstrated through measurement.

The dials are set by the organisation: not by the AI, and not by the vendors who supply it.

Human Gates: where the dials sit

A Human Gate is a deliberate checkpoint in an agentic workflow, and each one is governed by a Risk Dial. Some steps have no dial at all because they are always led by people; others start at Red and can move as evidence accumulates; a small number stay at Red permanently. The following shows the typical positions across an IDD delivery flow.

StepStarting positionHow it can move
User needs discovery (UCD)Always human-led, no dialDoes not move: people decide what to build
Domain modelling (DDD)Always human-led, no dialDoes not move: people define the business
Intent Specification and AI-generated testsRed: every test validatedAmber for coverage spot-checks, Green for low-risk work
AI-generated BDD scenariosRed: stakeholders validate all scenariosAmber for summary review, Green for notification only
Code reviewRed for security, payments, personal data and core infrastructureThose areas stay Red; lower-risk changes can move to Amber or Green
Security review of production systemsRedRarely moves
Agentic exploratory testingRed: all findings reviewedAmber for critical findings only, Green for a summary
User validation (UCD)Red: user testing on every releaseAmber: user testing on major features
DeploymentRed: every production deploy approvedAmber for approval by notification, Green with intervention and rollback available

Between those gates, implementation itself carries no gate, because the validation defined in the specification follows immediately and is fully automated.

Moving a dial

A dial moves on evidence, and the evidence is Intent Fidelity. Full review applies while fidelity is below 85% or during early implementations. Spot-checking becomes appropriate when fidelity is sustained between 85 and 95% across twenty or more implementations, and monitoring alone when it stays above 95% across fifty or more. Every move is reversible, and moving a dial back towards Red when conditions change is disciplined risk management rather than failure.

In multi-agent systems the dials can be applied at the level of the individual specialist agent, because different agents earn trust at different rates. An architecture agent might remain under full review while a test agent runs under monitoring and a security agent stays permanently at Red.

What never moves off Red

Some decisions remain under full human control regardless of maturity: strategic intent, ethical boundaries, security and compliance decisions, and final production approval. Agents do not get a vote on strategic direction or ethical judgement, and these permanent positions define one side of the Autonomy Boundary.

Where it comes from

Risk Dials were introduced in Human Gates in Agentic Flows for Enterprise AI Control, published on 1 February 2026, and extended to multi-agent systems in IDD via Multi-Agent Systems. They form Chapter 5 of the ebook.

Frequently asked questions

Where should a team set its dials when it starts?Everything that has a dial should start at Red. Full review at the outset costs some speed, but it is how a team calibrates where AI performs reliably and builds the fidelity record it will later need to justify any reduction in oversight.
Can a dial move back to Red once it has moved?Yes. Dials are designed to be reversible. A change of model, a new kind of work or a fall in measured fidelity are all good reasons to tighten control again, and doing so is a sign of mature governance rather than a setback.
Who decides where each dial is set?The organisation decides, through the people accountable for the work. Neither the AI system nor its vendor sets the level of autonomy, and dial positions should be recorded and reviewed in the same way as any other governance decision.
Are Risk Dials the same as human-in-the-loop?They are a structured form of it. Rather than a single yes-or-no choice about keeping people in the loop, Risk Dials give each step its own level of oversight, tied to measured evidence and adjustable over time.