# 13. Human oversight of AI decisions > **Template.** Part of the open-source [AI App Policy Kit](https://www.puglieseweb.com/ai-policy-kit), version 1.0. Replace everything in [square brackets], delete what does not apply, and have the person who runs the business approve it. A starting point, not legal advice. | | | |---|---| | **Business** | [Business name] | | **Document owner** | [Name and role] | | **Approved by, and date** | [Name], [date] | | **Next review** | [Date, at most 12 months on] | ## Why this matters AI systems are wrong some of the time, and they are wrong fluently. Where a wrong answer costs a customer money, an opportunity or their trust, a person must be able to check it, stop it and put it right. UK data protection law also gives people rights when a significant decision about them is made by automated means alone. ## Who it covers Every AI feature at [Business name] that decides, recommends or acts: chat assistants, scoring, screening, pricing, automated replies and AI agents. ## The rules 1. **Sort every AI use into a level** using the table below, and record it in the register. 2. **Significant decisions about a person are made by a person.** Refusing a service, a job or credit, setting an individual's price, closing an account or resolving a complaint is never left to AI alone. The person deciding has the authority and the information to disagree with the AI, and the time to do so. 3. **People are told.** Customers know when they are dealing with an AI system, and how to reach a person. 4. **There is always a way to a person.** Every AI conversation offers a hand-over, and it is answered within [one working day]. 5. **Results can be challenged.** A person affected by an AI-assisted decision can ask for an explanation and a review by someone who was not involved. Requests are logged. 6. **Actions have limits.** An AI agent that can send, book, pay, refund or delete has a ceiling on each (amount, number per day, who it can contact). Anything above the ceiling waits for a person's approval. 7. **There is an off switch.** [Name] can turn each AI feature off within minutes, and the business can carry on without it. The steps are written down. 8. **Check the output regularly.** Each [month], someone reviews a sample of AI outputs for accuracy, tone and fairness, and records what they find. 9. **Watch for unfair patterns.** Where AI affects how people are treated, compare outcomes across groups of customers at least every 12 months. 10. **Mistakes are corrected and learned from.** A wrong output that reached a customer is fixed with the customer, logged, and used to improve the prompt, the data or the limits. ## Levels of oversight | Level | What the AI does | Oversight needed | |---|---|---| | 1. Helps staff | Drafts, summarises, suggests. A member of staff decides what to use. | Staff check the output (document 01). | | 2. Talks to customers | Answers questions, with no effect on what the customer gets. | Customers are told. Hand-over to a person. Monthly sample. | | 3. Acts within limits | Books, sends, updates records, gives small refunds. | Written limits. Every action logged. Approval above the limit. Off switch tested. | | 4. Affects a person's rights or money | Screening, scoring, pricing, eligibility. | A person decides every case. Risk assessment first (document 09). Fairness check. | ## Register of AI uses | AI feature | Level | Limits | Who reviews, how often | Off switch held by | Last reviewed | |---|---|---|---|---|---| | [Website assistant] | [2] | [Cannot quote prices] | [Name, monthly] | [Name] | [Date] | | [ ] | [ ] | [ ] | [ ] | [ ] | [ ] | ## Review record | Date | Feature | Sample size | Problems found | Action | |---|---|---|---|---| | [ ] | [ ] | [ ] | [ ] | [ ] | ## Checklist - [ ] Every AI feature is in the register with a level. - [ ] No level 4 decision is made without a person. - [ ] The off switch for each feature was tested in the last 12 months. - [ ] The monthly sample reviews are recorded. - [ ] Challenges and their outcomes are logged. ## References - ICO, *Rights related to automated decision-making including profiling*. - ICO and The Alan Turing Institute, *Explaining decisions made with AI*. --- AI App Policy Kit, version 1.0. © 2026 PuglieseWeb LTD, licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/): use it, change it and share it, with credit. The whole kit, and a free Software Health Check of your app against it: