What leaders should never delegate to AI

AI can now draft the redundancy letter, rank the job applicants, price the contract, and suggest which supplier to drop. It can do all of these things quickly and often well. The danger is quiet and specific: because AI clearly can contribute to these decisions, it becomes tempting to let it make them. Capability slides into authority without anyone deciding that it should.

The useful question for a leader is not whether AI can help with a decision. It nearly always can. The question is whether the final responsibility for that decision can safely be separated from the person who is answerable for it. Often it cannot, and knowing which decisions those are is one of the more important judgements a leader now has to make.

Responsibility does not transfer to a model

Start from a fact that does not change with the technology: a model cannot be accountable. When an AI-assisted decision goes wrong, the responsibility sits exactly where it sat before, with the person and the organisation that made it. The OECD's AI Principles keep accountability and human oversight at the centre of trustworthy AI for this reason, and UK law draws a hard line in the same place. Where a decision is made solely by automated means and has a legal or similarly significant effect on someone, Article 22 of the UK GDPR restricts it and the Information Commissioner's Office expects human involvement to be active, not a rubber stamp.

So the line is not drawn around what AI can touch. It is drawn around where responsibility must stay attached to a human who can answer for the outcome. The five questions below are how we find that line.

The HUMAN test

Run a decision through these five. The more of them it triggers, the more firmly it should stay with a person, with AI informing the call but never closing it.

High consequence. What happens if this goes wrong? A decision that can seriously harm a person, the business, or a client is one to keep close. The cost of a confident error is the first thing to weigh.

Unclear or contested. Is there a single right answer, or does this turn on judgement, values and competing interests? AI is at its weakest where the question is genuinely arguable, and at its most dangerous when it makes a contested call sound settled.

Morally or relationally significant. Does this touch how you treat people, or a relationship that matters? Telling someone they have lost their job, or that a long-standing client is being dropped, is not a task to optimise. It carries meaning a model does not feel.

Accountable to law, customers or employees. Will you have to stand behind this to a regulator, a court, a customer or your own staff? If you would have to defend it, you must be the one who made it.

Not readily reversible. Can a mistake be quietly undone, or does it land and stay? The harder a decision is to take back, the less it should be handed over. Reversible calls are where automation is safe; irreversible ones are where judgement earns its keep.

AI can inform the decision. It cannot hold the responsibility.

What this looks like in practice

Consider a firm using AI across its hiring. Using it to draft the advert, summarise applications, or suggest interview questions is sensible: low stakes, reversible, easy to check. Letting it decide who is rejected is a different thing entirely. That decision is high-consequence, contested, significant to a person, accountable to employment law, and not easily reversed. It triggers all five, so it stays with a human, even though the AI could produce a ranking in seconds. The NIST AI Risk Management Framework makes the same point in general terms: manage a system in proportion to its consequences and context.

The pattern repeats wherever the stakes are real. A professional services firm can let AI prepare and research a piece of client advice, but the advice the client acts on, and pays for, is given by a named person who stands behind it. An accountancy practice can automate reminders and first-pass checks, but the sign-off on a filing is a human's. (Illustrative examples, not specific firms.) In each case AI does a great deal of the work and none of the deciding.

The honest limits

Two things this is not. First, it is not an argument to keep AI away from important work. The opposite, in fact: AI can strengthen a high-stakes decision enormously by widening the options, surfacing what a tired human missed, and pressure-testing the reasoning. The line falls on who decides and is answerable, not on how much AI contributes up to that point. Anthropic's own guidance on building agents makes the related point that handing a system autonomy means handing over judgement, which is precisely what these decisions cannot spare.

Second, the boundary is not fixed for all time, and it is not the same in every business. A decision that is high-consequence in a hospital is trivial in a marketing team. The test is a way to think, not a list to copy. Run your own decisions through it rather than borrowing someone else's answers.

What to do next

List the recurring decisions in your area and run each through the five questions. Mark the ones that trigger two or more as decisions that keep a human in the chair, and be explicit that AI may inform them but never finalise them. For the rest, let AI help freely. The clarity is worth more than it sounds: most teams have never said out loud which decisions are theirs to keep, and the ambiguity is where the accidents happen. Make the kept decisions visible to the people doing the work, so everyone knows where AI's help is welcome and where it stops, and write down who owns each one. A boundary that lives only in a leader's head is not a boundary the team can actually hold to.

The tool

To make that sortable in an afternoon, we have built the HUMAN Delegation Checklist: a one-page checklist that runs a decision through the five questions and lands it in a green, amber or red zone, with plain guidance on what AI may and may not do at each level.

Download the HUMAN Delegation Checklist (PDF)

Drawing these lines well is part of the governance an AI Advisory engagement leaves in place. It builds directly on deciding how much to hand to AI in the first place and on giving any AI agent a defined job, and it is the judgement our AI lessons for leaders are designed to sharpen.

Sources and further reading

• OECD AI Principles. Adopted 2019, updated 2024. Source for accountability and human oversight as core commitments of trustworthy AI.
• ICO, Rights related to automated decision-making including profiling. UK regulator. Source for the restriction on solely automated decisions with legal or significant effects.
• NIST AI Risk Management Framework. Independent US standards body. Source for managing systems in proportion to consequence and context.
• Anthropic, Building Effective Agents, December 2024. Vendor guidance from a model developer. Source for autonomy meaning the delegation of judgement.