How should AI decision rights be allocated?

By naming the roles separately, the way any decision has always had to be allocated. Who recommends. Who supplies input. Who has to agree. Who decides. Who executes.

Share this line as a card

AI can now perform parts of several of those. It cannot hold the decide role, because deciding carries answerability, and a system cannot be answerable. So the structure of the exercise is unchanged and the occupants of the boxes have moved.

Separating the work from the right#

The confusion this question usually contains is between doing the work of deciding and holding the right to decide. They were always different, and they were easy to conflate while the same person did both.

Gathering the evidence, weighing the options and drafting the recommendation is work. It can be delegated, to a junior, to a supplier, or now to a system. The right to decide is a position in an accountability structure, and delegating the work has never transferred it.

Which makes the AI-specific part narrower than it looks. Most of what these systems do is input and recommendation, and organisations have been allocating those roles to other parties for as long as organisations have existed.

The allocation that goes unmade#

The practical failure is not a wrong allocation but an absent one.

For each decision the organisation should be able to state whether the system recommends, decides within stated limits, or executes, and what those limits are. In most deployments none of this is written down. The system produces an output, the output is accepted, and the decide role has been given away by default rather than by choice.

Nobody made that decision, which is the whole problem: an allocation arrived at through drift cannot be reviewed, because there is no record of it having been made.

This research treats the mapping exercise itself at the delegation boundary map, and the point that the resulting artefact is a record of decisions rather than a policy document applies here too.

What the regulation already allocates#

The EU AI Act distinguishes the provider, who develops a system and places it on the market, from the deployer, who uses it under their own authority, and places obligations on both. Graded entry.

That distinction does real work, because it forecloses the most convenient answer. An organisation using a system cannot locate responsibility entirely with whoever built it. Using it under your authority is what makes you a deployer, and deployer obligations attach to that.

Article 14 then says what the person given oversight must be enabled to do, which is a decision-rights allocation written as a design requirement: they must be able to disregard, override or reverse the output, and to stop the system.

Accountability that is real and accountability that is nominal#

Accountability can only attach to a person or an organisation, so the question of who is accountable when AI gets something wrong has a short answer. The useful question is whether that accountability is fair, and often it is not.

Assigning responsibility to someone who could not have evaluated the output produces a name to blame rather than a control. It satisfies an audit and changes nothing about the failure rate, because the person named was never in a position to prevent anything.

The test is the one the regulation uses in its single specific provision: competence, training and authority, together. Any one of the three missing and the accountability is nominal. Treated at length at who can override an AI system and who supervises work they cannot do.

What agents change#

Chains of agents complicate the tracing rather than the principle. Whoever deployed the chain deployed it.

What genuinely changes is speed. A mistake can propagate through several steps before any human reviews any of them, which means the allocation cannot be made during operation and has to exist beforehand. An agent architecture without a decided allocation is not an ungoverned decision; it is a governance decision made by omission.

Perrow's coupling argument applies directly. Normal Accidents (1984). Tighter coupling gives less warning between a fault and its consequences, and agents acting on each other's outputs is a description of tight coupling.

The last time technology moved these rights, the direction depended on which technology#

There is one measurement worth holding against all of the above, and it predates generative systems by a decade.

Nicholas Bloom, Luis Garicano, Raffaella Sadun and John Van Reenen built an original dataset of firms in the United States and seven European countries and measured three things: the autonomy of production workers, the autonomy of plant managers, and spans of control. They then separated the two halves of what is usually reported as one variable. Graded entry.

The two halves pulled in opposite directions. Technologies that lower the cost of knowing something, enterprise resource planning for plant managers and CAD/CAM for production workers, went with more autonomy and wider spans of control. Technologies that lower the cost of asking somebody else, data networks, went with less autonomy for workers and plant managers alike. The result survives their instrument, which uses differences in telecommunication costs across countries arising from different regulatory regimes.

Read that as a warning about drift rather than as a prediction. Decision rights move when the cost of information moves, whether or not anyone allocates them, and the direction follows what the technology makes cheaper. A system that puts more of the relevant knowledge in front of the person doing the work pushes the right to decide towards that person. A system that makes referring upwards cheaper pulls it away.

Generative systems do both at once. They put more knowledge in front of the junior analyst and they make it cheaper to obtain an answer from somewhere other than the analyst's own head, so the drift has no direction anyone can predict from the technology alone. That is the case for writing the allocation down instead of watching it settle.

Five roles offered for clarity, not because they were tested#

The five-role framing is a widely used organisational-design convention rather than a validated construct, and several proprietary versions of it exist. It is offered because it is clear, not because a study shows it produces better decisions.

The EU AI Act is binding law and not evidence about outcomes, and its oversight provisions do not apply until 2 December 2027 as a longstop rather than a start date. Nothing here shows that organisations allocating rights explicitly perform better than those that do not.

Bloom and colleagues measured manufacturing plants and technologies of the 1990s and 2000s. Their results are associations between technology use and organisational form, with one instrument doing the causal work, and no part of the study concerns AI. The section above is read at source from the authors' own summary of the published paper rather than from its tables, so it carries no coefficients and none should be attributed to it.

The fairness argument is a normative position rather than a finding. Someone could reasonably hold that accepting accountability for outputs one cannot fully evaluate is simply what senior roles have always involved, and that view is not refuted by anything on this page.

Key sources

• European Union (2024). Regulation (EU) 2024/1689, Article 14: Human Oversight. Graded entry.
• Perrow, C. (1984). Normal Accidents: Living with High-Risk Technologies. Basic Books. In the essential works.
• Weick, K. E. and Sutcliffe, K. M. (2001). Managing the Unexpected. Jossey-Bass. In the essential works.
• National Institute of Standards and Technology (2023). AI Risk Management Framework Playbook, MANAGE 2.4. Graded entry.
• Bloom, N., Garicano, L., Sadun, R. and Van Reenen, J. (2014). The Distinct Effects of Information Technology and Communication Technology on Firm Organization. Management Science, 60(12), 2859-2885. Graded entry.

Related SuperSkills research#

On mapping the boundary, the delegation boundary map. On the authority to act, who can override an AI system. On agents acting independently, AI agents and human judgement and should I let an agent act on my behalf. On verification ownership, who owns verification. On withdrawing a system, deployment is not a ratchet.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The separation of decision work from decision rights is standard organisational design and is not a SuperSkills coinage.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.