Human at the Start

Every board has someone who says it: we keep a human in the loop. It is the most reassuring sentence in corporate governance, and on its own it protects nobody. The real risk of AI is not automation, which is visible, budgeted and argued over. It is accountability: a decision passing through a model and a chain of people until no one can honestly say they made it. AI launders accountability. Responsibility goes in one end, plausibility comes out the other, and the trail between them is gone. The answer is not more oversight in the middle. It is a named person at the start, who frames the decision, and a named person at the end, who owns it. There are three places to put a human in relation to a machine's decision, and the one everybody names is the weakest.

The three positions

Human at the Start (HATS). Before a model is pointed at a problem, one named person defines what a good outcome looks like, what the tool may conclude, and what would cause them to override it. That is where the decision is made. Everything after it is administration, however senior the person performing it. This is the strongest position, because it puts judgement where judgement actually lives: in the framing.

Human in the Loop. This is the phrase boards reach for, and the weakest of the three. European regulators dealt with it years ago. Guidance endorsed by the European Data Protection Board holds that a controller cannot escape Article 22 of the GDPR by manufacturing human involvement, and that oversight must come from someone with the authority and competence to change the decision. In SCHUFA, the Court of Justice held that producing an automated credit score can itself amount to the automated decision, where the lender draws strongly on it. A person placed inside the process who has no power to change the outcome is a witness to it, not a decision-maker.

Human at the End (HATE). This is the position most organisations think they already hold, and it is not a signature. It is an owner: someone who can explain the outcome without reference to the tool, and who had the standing to refuse it. If you want to automate a consequential decision, you need this position and the first one. The loop between them will not save you.

How accountability disappears

Consider a real case. Frances Walter was an eighty-five-year-old in Wisconsin with a shattered shoulder and an allergy to most pain medication. An algorithm called nH Predict, matching her against a database of six million patients, estimated she would be ready to leave her nursing home in 16.6 days. A reviewer entered that estimate in her file. A medical director later cited it in finding she no longer met her insurer's coverage rules, and payment stopped on the seventeenth day, while her notes recorded her pain at the top of the scale. A judge later called the denial, at best, speculative. There were two humans in that loop. No shoulder heals to a tenth of a day, but precision reads as authority, and everyone in the chain treated the estimate as though someone had thought hard about it. The algorithm did not decide; it produced an estimate. The reviewer did not decide; she recorded it. The director applied criteria to an assessment already made for him. There is no villain in that sequence, and no decision-maker in it either.

Responsibility thins at each handoff. A model's output arrives finished, and disagreeing with it feels less like judgement than accusation, so the analyst accepts it, the manager approves the analyst, the director approves the manager, and the board notes the outcome. Everyone acted reasonably; the aggregate is unreasonable; and the audit trail shows four approvals and no objections, which is exactly what it was designed to show. Then the incentives arrive: in that case, managers were set a target of keeping stays within three per cent of the algorithm's projection, later tightened to one. Deference was simply the cheaper of the two available behaviours, and it was recorded in the file as agreement. Across seven years of research in more than 200 organisations, I have asked leaders to walk me through a decision their AI touched, and I have never once been given a name without a pause coming first. The pause is the finding.

Why the loop is where oversight is weakest

The case for the start and the end over the loop is also what the evidence on late oversight shows. Parasuraman and Manzey, reviewing decades of work across aviation, medicine and the military, found that people tend to under-question confident automated output, that this appears in experts as much as novices, that it cannot be trained away, and that it worsens under load. A human placed in the middle, under time pressure and, as in the case above, measured on how closely they agree with the machine, is placed exactly where oversight is least reliable. The loop feels like control and delivers very little of it.

The test that restores accountability

Four things, then, before a model touches a consequential decision.

• One named individual owns it, in writing. Not the model, not the function, not the committee.
• That person can explain it without reference to the tool. If the only justification is that the system said so, there is none.
• That person has the authority to overrule it, and the time, and is not marked down for using either.
• If no name can be attached, the decision is not ready to be automated.

The same word runs through the law: authority. The EU AI Act requires deployers of high-risk systems to give oversight to people with the competence, training and authority to act on it. The data protection test for meaningful human involvement turns on authority. Under the UK Senior Managers and Certification Regime, British financial services has run the principle as law for a decade: every senior manager holds a Statement of Responsibilities naming what is theirs, and any delegation of it must go to an appropriate person whom they then oversee. A model is not an appropriate person, and neither is a signatory who cannot say no. For directors this is not housekeeping: section 174 of the Companies Act requires reasonable care, skill and diligence, personally, and that duty does not transfer to a vendor when the reasoning moves into software.

The bottom line

None of this is free. Naming an owner slows decisions down, and it will cause capable people to refuse to sign things they ought to have signed. Some refusals will be wrong, and they will cost real money. Anyone selling named accountability as a free good is not being straight. But somewhere in your organisation is a decision from this quarter that nobody can honestly claim. It was probably fine; most of them are. An organisation that cannot name who made a decision, though, has not automated that decision. It has abandoned it. Human at the Start, with a named owner at the end, is how you take it back.

Key sources

• Hirji, R. (2026). Why the Real AI Risk is Not Automation, but Accountability Gaps in Leadership Decisions. The European Business Review.
• STAT News (2023). Denied by AI: how Medicare Advantage plans use algorithms to cut off care.
• Court of Justice of the EU (2023). SCHUFA Holding, Case C-634/21, on automated decisions and Article 22 GDPR.
• Regulation (EU) 2024/1689 (the EU AI Act), on human oversight of high-risk systems.
• Parasuraman, R. and Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation. Human Factors, 52(3).

Related SuperSkills research

Human at the Start connects to AI and human judgement, AI and critical thinking, the Augmented Mindset, drift versus design and capability debt.

About this research

Rahim Hirji is the author of SuperSkills: The Seven Human Skills for the Age of AI (Kogan Page, 2026) and the founder of The SuperSkills Intelligence Company. This work draws on research across more than 200 organisations in 30 countries over seven years, and the argument here was set out in The European Business Review. Human at the Start and Human at the End are part of the SuperSkills lexicon; the case, cases and legal instruments are cited to their sources and kept separate from the framework, which is the author's. This is a living reference, reviewed and updated as significant new evidence appears.