Institute Tool

Your policy may say a human remains accountable. But can that person understand, challenge and stop the AI output before it is used to shortlist a candidate, assess an employee, approve a customer request or recommend a major organisational action?

A reviewer may appear in the process while lacking the time, evidence, expertise or authority required to disagree with the system. The approval remains human in name but becomes automatic in practice.

The Human Accountability Boundary helps an organisation decide what AI may do, what information it may use, what a qualified person must verify and who can reject, override, stop or escalate the output.

By Rob Adesanya · The Rob Adesanya Institute

The Oversight Test

Is the human reviewer genuinely able to disagree?

Human oversight is weak when a person can approve the result but cannot inspect the evidence, identify an error or stop the process before the output is used.

The problem is not solved by adding a name to an approval box. The reviewer needs enough information, competence, time and authority to exercise independent judgement.

Oversight may be weaker than it appears when

  • The reviewer sees only the recommendation, not the underlying evidence.
  • The system produces more cases than a person can meaningfully examine.
  • The reviewer lacks the expertise required to identify an error.
  • Challenging the system creates delay or additional work.
  • Nobody knows who can override the output.
  • An employee believes rejecting the recommendation will damage their credibility.
  • The system's confident language is mistaken for reliable evidence.
  • Approval happens after the action has effectively become irreversible.
  • The organisation cannot explain the result to the person affected.
  • Responsibility is divided between the user, manager, vendor and technical team until nobody clearly owns it.

AI reduces the cost of choices not the cost of choosing badly

AI can generate alternatives at a speed that would once have required a large team and several weeks. It can compare scenarios, summarise evidence, expose patterns and draft recommendations. This changes the economics of analysis. It does not remove the cost of a poor decision.

That distinction matters because organisations may confuse the availability of more options with an improvement in judgement. More options can widen a leader's view. They can also create false confidence, conceal weak assumptions and make responsibility easier to disperse.

Felipe A. Csaszar's Harvard Business Review article on AI and strategic decision-making examines how AI expands the possibilities organisations can consider. When options become cheaper to produce, the scarce capability is deciding which deserve belief, which risks are acceptable and who will answer for the result.

The hidden transfer of responsibility

Responsibility rarely disappears openly. It moves through ordinary language. A manager says that the system identified the strongest candidate. A project team says that the model recommended the investment. A leader presents an AI-generated forecast as though the fluency of the explanation proves the quality of the evidence.

These statements can make a human choice sound inevitable. The system produced an output. People still selected the data, accepted the criteria, decided how much scrutiny to apply and chose whether to act.

A policy may require human oversight, while the practical process gives the reviewer too little time, expertise or authority to challenge the machine. The approval remains human in name but becomes automatic in practice.

A boundary before a decision

The Rob Adesanya Institute uses the term Human Accountability Boundary for the point at which AI assistance must stop and accountable human judgement must take over. The boundary varies with the work. Drafting a routine internal summary does not require the same control as selecting employees for redundancy, approving credit, diagnosing illness or making a public allegation.

The boundary should be agreed before the output creates momentum. Leaders should first decide what the system may do, what must be checked and what it must never determine independently.

Six questions for leaders

  1. What decision or action is the AI system influencing?
  2. What information will the system receive, and is its use authorised?
  3. Who may benefit, lose or carry the consequences?
  4. How serious would an error be, and could it be reversed?
  5. What evidence and qualified expertise will verify the output?
  6. Who can approve, reject, override, stop or escalate the process?

Four decision levels

1

Assist

AI may generate options, summarise material, draft or organise information. A competent user checks the output before use.

2

Recommend

AI may compare options or propose a course of action using defined criteria. A named reviewer tests the evidence, assumptions and possible effects.

3

Human Decision

AI may provide analysis, but an authorised person makes, records and can explain the final decision.

4

Stop or Escalate

Pause the process when authority, evidence, information sensitivity, explainability or possible harm is unclear. Move the issue to the appropriate senior, legal, HR, risk, security or technical authority.

Decision rule: Move towards qualified human decision, stopping or escalation as consequences become more serious, errors become harder to reverse, information becomes more sensitive or accountability becomes less clear.

Applied case: AI-supported workforce restructuring

This fictional educational case concerns a financial-services organisation considering a restructuring. Leaders propose using AI to analyse role descriptions, workload data, organisational layers and cost scenarios. The system can process more possibilities than the project team could examine manually. The decision still affects employment, trust, operational continuity and the organisation's ability to explain its reasoning.

LevelApplication in this case
AssistOrganise role information, identify duplication, summarise consultation material and model alternative structures.
RecommendCompare scenarios against human-defined criteria. A qualified team tests the source data, assumptions, omissions and potential bias.
Human DecisionAuthorised leaders decide the purpose, criteria, affected roles and treatment of employees. AI must not independently determine who loses employment or communicate the final employment decision.
Stop or EscalatePause if sensitive data lacks authority, the output cannot be explained or challenged, protected characteristics may influence results, evidence conflicts or the action may breach legal or policy requirements.

Questions for the leadership team

  • Which parts of the work genuinely benefit from AI?
  • Which recommendation would be difficult to defend to an affected employee?
  • What information is missing from the dataset but visible to experienced managers?
  • Who could stop the process if the analysis appeared efficient but unfair?
  • What should be recorded before the decision is implemented?

The organisation must establish lawful authority, appropriate HR and legal involvement, suitable data controls and a fair process. Completing a Human Accountability Boundary does not prove that the proposed use is lawful, fair or safe.

Better options still require better judgement

AI can improve the range and speed of analysis. It can help a leader see an alternative that would otherwise have been missed. The final advantage does not come from generating the longest list of possibilities. It comes from knowing what to eliminate, what to test, what to escalate and what risk to accept.

AI reduces the cost of generating choices. It does not reduce the cost of choosing badly. Organisations that understand the difference can gain value from the technology without weakening the human accountability on which trust depends.

When to Set or Review the Boundary

Do it before the output creates momentum, cost and organisational dependence.

Before approval or purchase

Examine the intended task, information, decision authority and consequences before committing to a system.

Before a pilot begins

Define permitted use, verification, success measures, review dates and stopping conditions.

Before wider deployment

Reassess the boundary when a tool moves from experiment to ordinary working practice.

When people may be affected

Review the boundary before an output is used in employment, customer, professional or other important decisions.

When data or criteria change

Do not assume the old boundary remains suitable after the information, model, criteria or purpose changes.

After an error or complaint

Examine whether the reviewer could identify, challenge and stop the problem in time.

When ownership is unclear

Resolve responsibility between users, managers, technical teams, vendors and authorised decision-makers.

Leadership Decision Record

Make the boundary explicit before the work begins.

Use this worksheet for one real AI-supported task, recommendation or decision. Your entries remain in your browser and are not sent to the Institute.

This page does not transmit your entries to the Institute. Do not enter confidential, personal, legally privileged or security-sensitive information. Printing or saving creates a copy on your own device or chosen destination.

Important Boundary

Completing the worksheet does not authorise an AI use case.

The Human Accountability Boundary is not a legal-compliance checklist, data-protection impact assessment, cybersecurity review, technical model audit or independent assurance that a system is safe, fair or lawful.

It does not replace sector-specific governance or specialist professional advice, and it does not transfer accountability to Rob Adesanya or the Institute.

Pause and obtain specialist review when needed

  • Personal, confidential or protected information is involved.
  • The system may affect employment, access, safety, rights or significant financial outcomes.
  • The evidence, model behaviour or legal authority cannot be explained.
  • The reviewer lacks the competence, time or power to disagree.
  • A material change has invalidated the original boundary.
Sources and Context

Connect leadership judgement with established AI governance.

The boundary is an applied leadership tool within the Institute's Consequential Judgement framework. It complements formal governance, risk, technical and legal controls; it does not replace them.

Choose How to Apply the Boundary

Use it independently or examine a live organisational use case.

Use independently

Complete and print the Decision Record for one appropriately anonymised or non-sensitive use case.

Use the Decision Record

Examine one use case

Request a private AI Use-Case Lab to test purpose, information, verification, authority and stopping conditions.

Discuss a Use-Case Lab

Prepare leaders or teams

Use a leadership briefing, team workshop or modular programme to develop shared judgement and working boundaries.

Explore AI Adoption Training

Examine the underlying judgement

Use the Consequential Judgement framework to test evidence, stakeholders, alternatives and accepted consequences.

Explore Consequential Judgement

A human name on the approval box is not enough.

The person must be able to understand, challenge and stop the process before the AI output is used to make or support an important decision.