If you are putting AI into decisions about customers or staff, such as credit, eligibility, hiring or pricing, you have probably built a person into the process. Someone looks at the AI's output before it counts. That is your human in the loop.
Sooner or later someone will ask you to show it. It might be an auditor, a regulator, a large customer's assurance team, or a customer who wants to contest a decision. This post is about what you hand them.
The Short Answer
"A person is in the loop" is a claim. When someone asks, you need to show who checked, what they saw, what they decided and when.
The rules in the EU and the UK point the same way. The EU AI Act asks for human oversight by people with the competence, training and authority to do it. UK data protection law now turns on whether there was "meaningful human involvement" in a decision. A rubber stamp meets neither.
Most teams can say a person is in the loop. Fewer can show it, decision by decision.
What the EU AI Act Asks For
The EU AI Act sorts AI uses by risk. The oversight duties in this post apply to high-risk systems, and Annex III of the Act lists the uses that count. Examples include:
- Assessing the creditworthiness or credit score of individuals (fraud detection is excluded).
- Recruiting and selecting candidates.
- Decisions about promotion, termination and task allocation at work.
- Risk assessment and pricing for life and health insurance.
Not every AI use is high-risk, so check yours against the list.
What the system must allow. Article 14 of the AI Act(opens in new tab) says high-risk systems must be built so that people can oversee them effectively while they are in use. The people doing it must be able to understand what the system can and cannot do, and watch for over-reliance on its output (the Act calls this automation bias). They must be able to interpret the output, disregard or override it, and stop the system.
What you must do as the user. Article 26 covers deployers, the organisations that use a high-risk system. You must give the job of human oversight to people with the competence, training and authority to do it. You must use the system as its instructions say and monitor how it runs. You must keep the logs the system generates for at least six months. Where a high-risk system makes or helps make decisions about people, you must tell those people.
What the system must record. Article 12 says high-risk systems must automatically record events over their lifetime. That is the system's log.
When it applies. The high-risk rules were postponed, not cancelled. They were due to apply from 2 August 2026. Regulation (EU) 2026/1744 changed that. It was published on 24 July 2026 and came into force on 27 July 2026. High-risk systems listed in Annex III now apply from 2 December 2027. High-risk AI inside products covered by Annex I applies from 2 August 2028.
Does it reach UK firms? In two cases, yes. A UK company is in scope if it places an AI system on the EU market, or if the output of its AI system is used in the EU. A UK firm serving only UK customers is not.
The UK: "Meaningful Human Involvement"
In the UK, the rules that matter most for decisions about people sit in the UK General Data Protection Regulation (UK GDPR), and they changed this year.
Since 5 February 2026, new Articles 22A to 22D have replaced the old Article 22. They were brought in by section 80 of the Data (Use and Access) Act 2025(opens in new tab). Under Article 22A, a decision is solely automated where there is "no meaningful human involvement in the taking of the decision". Article 22C gives people the right to "obtain human intervention on the part of the controller" and to "contest such decisions".
So the question is not only whether a person was involved. It is whether that involvement was meaningful.
The Information Commissioner's Office (ICO), the UK data protection regulator, has said what it does not look like. Its guidance on individual rights in AI systems(opens in new tab) says "a decision does not fall outside the scope of Article 22 just because a human has 'rubber-stamped' it". It says "the human input needs to be meaningful", and that a reviewer's involvement must be "active and not just a token gesture".
Two caveats. The ICO marks that page as under review because of the 2025 Act, and its new guidance on automated decision-making is still in draft. The words "meaningful human involvement" are now in the law itself, but the ICO has not yet published a final view of what they mean under Article 22A.
The Financial Conduct Authority (FCA) has said it does not plan new AI-specific rules. In its responses to questions from its 2026 Annual Public Meeting(opens in new tab), published on 6 November 2026, it said it relies on existing frameworks such as the Consumer Duty. It added:
Authorised firms, advisers and senior managers remain responsible for outcomes where AI is used.
That answer was about advice and consumer-facing AI. It does not mean the FCA requires human review of AI decisions. It does mean the responsibility for the outcome stays with your firm.
In the SCHUFA case (C-634/21, 7 December 2023), the Court of Justice of the European Union (CJEU) found that a credit score could itself count as an automated decision under EU GDPR where the lender relied heavily on it. That held even though a person at the lender made the final call. The judgment(opens in new tab) is EU law and only persuasive in the UK. Our reading: a sign-off at the end may not help if the score in effect made the decision.
Why "A Human Was in the Loop" Isn't Proof
This section is our view, not the law.
The AI Act asks a high-risk system to log its own events, and asks you to keep those logs. It does not say you must log each human review. A system log shows what the AI did. It does not show that a competent person looked at the output, understood it, could have overridden it, and then decided.
It matters because a rubber stamp and a real review look the same in a system log. In both, the AI produced an output and a decision followed. The difference is everything in between: who looked, what they saw, and whether they could have said no. A log is not proof, which is the point we make on our page about the record.
Who tends to ask:
- An auditor, checking that the control you describe runs.
- A regulator, asking how you meet a duty such as Article 26 of the AI Act, or the safeguards in Article 22C of UK GDPR.
- A large customer's assurance team, before they sign or renew.
- A customer who contests a decision, and asks for a person to look at it again.
When that question comes, a team that kept a record of each check has a straight answer. A team with only a system log has to rebuild the story from memory, emails and whoever is still around.
What a Record of a Human Check Should Contain
This is good practice, not a legal checklist. For each decision a person reviews, the record should show:
- What the reviewer saw. The input and the AI's output, as they were at the time.
- Who the reviewer was. Their name and role, and why they were competent to review it, such as the training they had.
- What they decided. Whether they accepted, changed or rejected the AI's output, and why.
- When. The date and time of the review.
- What happened next. The outcome that reached the customer or member of staff.
- Whether it changed later. A record that shows if any of the above was edited after the event.
Two questions test whether the record works. Could someone outside your team read one record and see that the review was real? And would a rubber stamp look any different from a real review? If the answer to the second is no, the record is not doing its job.
If your reviewers can override the AI, the record should show the times they did. A review step that never changes anything is worth a second look.
Where to Start
- Pick one decision. The one with the most effect on people, or the likeliest to be asked about.
- Write down who reviews it and what they must see. Name the role, the training it needs, and the information the reviewer has in front of them.
- Record each review. Use the list above, from the next decision onwards.
- Test the handover. Ask whether you could hand the record to an auditor today. If not, fix what is missing.
- Review it before your first audit, not after. For Annex III uses in the EU, 2 December 2027 is the date to plan against.
How AUDITSU Helps
Your team owns its decisions and the oversight of them. Our part is the evidence.
Most tools give you a log. A record says who allowed each decision, on what evidence and when.
Through AI Harness Engineering, we build approval gates with named people at the steps that carry consequence. We also build checks against your rules that run before an action takes effect. The rules are yours: written in your words, versioned, and changed by your team with a sign-off recorded against each change.
The software does not decide whether you comply. The check produces the evidence, and a named person judges. In our accessibility product today, nothing a scanner finds reaches your record until a person on your team accepts it, and every verdict is recorded against the person who made it.
If you have been asked to prove something, bring the request. We will tell you what you could show and what it would take.