Skip to content

The decision chain

The decision chain is the division of work in a consequential judgement: a platform assists, a consultant authors the recommendation, the accountable leader decides, and the observed outcome scores the call. Each step has a different owner, so responsibility for the decision stays with a person and the result can later be compared with the forecast.

Platform assists, consultant authors, leader decides, reality scores: each step has a separate owner, and the last word goes to the outcome.

Why it matters when the plan changes

When software contributes to a decision about people, responsibility tends to blur. The system produced the score, the consultant passed it on, the leader relied on it, and nobody quite owns the result. Research on human and AI combinations shows the cost of leaving roles undefined: a meta-analysis of 106 experimental studies found combinations on average did worse than the better of humans or AI alone, with losses in decision tasks.

The tension is between efficiency and authority. The more the platform does, the more tempting it is to let its output stand. Meaningful oversight requires information, competence, authority and time to challenge the system, which is also the direction of the EU AI Act for high-risk AI. The chain keeps a human author and a human decider, and lets the outcome, not the argument, settle who was right.

In practice

After an acquisition, the new owner asks whether the target's leadership team can carry the integration plan. The platform organises assessment evidence and documents. A consultant writes the forecast, including what argues against it. The chief executive decides to split one role and records why. At the agreed review, the outcome is compared with the forecast and the next call is adjusted.

Evidence

What it cannot tell you

The decision chain is a working name for a principle, and neither the chain nor the method around it has yet been validated against outcomes. Legal requirements for human oversight depend on the specific use and jurisdiction, so the chain describes intended roles rather than a legal position.

How Atlas reads it

Platform assists, consultant authors, leader decides, reality scores. Each role should see what helps them act, while the person remains protected and the final decision remains challengeable. Meaningful oversight requires information, competence, authority and time to challenge the system, and the legal requirements depend on the specific use and jurisdiction. The name is a working one.

Questions

Because the person who writes a recommendation and the person accountable for acting on it face different pressures. Keeping them separate means the recommendation can include evidence against itself without softening, and the decision-maker can reject it without having to rewrite it. Each role stays answerable for its own step.

That the outcome, not the quality of the argument, settles whether the call was right. The forecast names what will test it and when; at that point the observed result is compared with what was expected, and the comparison changes the evidence, the confidence and the next recommendation.

It reflects their logic without being a legal claim. Article 14 of the EU AI Act requires that high-risk AI systems can be effectively overseen by natural persons while in use. Whether a particular use falls under those rules depends on the use and the jurisdiction, and has to be assessed case by case.

Because accuracy on average is not accountability for a specific person's situation. A 2024 meta-analysis of 106 studies also found human-AI combinations underperforming the better of the two in decision tasks, which argues for defining each role carefully rather than removing the human or adding one without a clear task.

No. It is a working name for a principle the worldview shows without a name of its own, and the final name is still to be decided. The principle itself, a separate owner for assisting, authoring, deciding and scoring, is the part that matters for how a decision is run.