Keep These Decisions Human: A Practical Map of Where AI Doesn't Belong

Most AI advice tells you where to start. This is the other list: the decisions you should deliberately keep in human hands, written down before someone automates them by accident.

That last part matters. Almost nobody decides to hand a firing decision to a model. What happens instead is that a screening tool ranks candidates, a manager reads the ranking, and six months later nobody can explain why the shortlist looked the way it did. The boundary moved without a meeting.

The test: who answers for it when it goes wrong

Forget the technology for a moment and ask three questions about the decision itself.

Can you reverse it cheaply? A misrouted support ticket costs minutes. A rejected mortgage application, a withdrawn job offer or a terminated supplier contract costs far more, and the damage lands on someone outside your company.

Will you have to explain it to a person who is angry? Not a dashboard, not a board slide — a named individual who wants to know why. "The system scored you at 0.31" is not an explanation, and everyone in the room knows it.

Does the decision create the rules for other decisions? Setting a price floor, defining what counts as an acceptable customer, choosing which metric the sales team is paid on — these decisions propagate. Get one wrong and you have not made one mistake, you have made a policy.

If a decision fails two of those three, keep it human on purpose and say so in writing.

Four places to draw the line

Decisions about people. Hiring, promotion, performance ratings, discipline, redundancy. In the EU this is not just judgment, it is regulation: employment decisions sit in the high-risk category of the AI Act, which means documentation, human oversight and the ability to explain an outcome. Beyond compliance, there is a simpler reason. A manager who cannot defend a decision in their own words has lost the authority to make it.

Decisions that end a relationship. Closing an account, denying a claim, cutting off a supplier, refusing a refund above a threshold. Automation is excellent at spotting the candidates and terrible at knowing which of them is the customer who has been with you for eleven years and just had a bad quarter.

Decisions you will have to defend to a regulator or a court. Credit, insurance, medical triage, anything touching safety. The standard is not "was the model accurate" but "can you reconstruct, months later, what information was in front of whom." Build the reconstruction step first or don't automate at all.

Decisions about what to optimise. A model will pursue the target you give it with more discipline than any team you have ever managed. That is exactly why choosing the target is your job. Ask a system to maximise resolved tickets and it will close them faster, including the ones that needed a conversation.

The rubber stamp problem

Most companies think they have solved this with a human review step. Usually they have not. A review step is theatre if any of these is true:

• The reviewer approves more than 95% of what reaches them.
• Median review time is under thirty seconds for a decision that would take a person five minutes to think through properly.
• The reviewer sees a recommendation but not the evidence behind it, or cannot see the cases the system filtered out before they arrived.
• Nobody is measured on the quality of overrides, only on throughput.
• Reversing the recommendation requires more clicks, more justification or more explaining than accepting it.

Fix the incentives or drop the pretence. A rubber stamp is worse than full automation, because it gives you the liability of a human decision with none of the judgment.

Where the machine earns its place

None of this argues for keeping AI away from serious work. It argues for putting it on the right side of the line. Let it assemble the file: pull the contract history, summarise eighteen months of correspondence, flag the three clauses that differ from your standard terms, lay out what similar cases were decided and when. That is the heavy lifting, and it is genuinely hard for a person to do well under time pressure.

Then let a human make the call, with their name on it.

The companies getting the most out of AI are not the ones automating the most decisions. They are the ones whose people now walk into hard decisions prepared, instead of walking into them at 6pm with a half-read file. That is the split we build for at Hercules A, and it is worth deciding on paper — which decisions stay yours — before the tools decide it for you.