The question that catches businesses out is never ‘are you allowed to use AI’. It is ‘show me what happened’. A client queries an invoice, a candidate asks why they were rejected, an insurer asks how a figure was arrived at, and somebody has to reconstruct a decision made months ago by a person who has since left, using a tool that keeps no record. That is the gap this piece is about, and closing it is far less work than most people fear.
Compliance is demonstrating that your AI use meets duties you already have: UK GDPR and the DPA 2018 as amended by the Data (Use and Access) Act 2025, the Equality Act 2010, your regulator’s rules and your client contracts. There is no separate AI regime to comply with in the UK. There is your existing regime, plus evidence.
Auditing is checking that what you wrote down is what people actually do. Twice a year, internally, four questions: what tools are in use, what data went in, what output went out, and where is the human sign-off.
Explainability (XAI) is being able to say why an output happened, in terms the affected person accepts. You cannot explain the model. You can explain the process, and for business purposes that is what counts.
Compliance: your existing duties, applied
I want to clear up a confusion that costs money. Businesses keep asking me which AI law they need to comply with, as though there were a single statute waiting to be satisfied. There is not, in the UK. What there is instead is a stack of existing obligations that do not pause because you started using a chatbot.
| Obligation | What it asks of your AI use |
|---|---|
| UK GDPR and DPA 2018, as amended by the Data (Use and Access) Act 2025 | A lawful basis for the processing, data minimisation, transparency to the people involved, and a risk assessment where processing is likely to be high risk. Additional protections apply to solely automated decisions with significant effects. |
| Equality Act 2010 | No discrimination, direct or indirect, in anything that screens, scores or ranks people. Including where the discrimination is a side effect the tool inferred rather than one you intended. |
| Sector regulation (SRA, FCA, ICAEW, RICS, CQC and others) | Competence, supervision, client care, and record-keeping. Regulators have generally been clear that using a tool does not transfer professional responsibility to it. |
| Client contracts and NDAs | Confidentiality obligations that pasting a document into a third-party tool may breach, depending on the terms and the tool. |
| Employment law | Fair process in monitoring, performance management and dismissal, none of which becomes easier to defend if the reasoning was generated. |
| ISO/IEC 42001 (voluntary) | A structure for an AI management system, and something certifiable if clients start asking. Worth it mainly if you already run ISO 9001 or 27001. |
Two points. First, none of this is exotic. If you have data protection in reasonable order, your AI compliance is mostly a matter of extending what you already do. Second, and this is where the work sits, all of these are evidential. It is not enough to have been careful. You have to be able to show that you were, at a point in time, using records you made then rather than a reconstruction you assemble now.
Compliance you cannot evidence is indistinguishable from non-compliance you got away with.
The records that make this survivable
Here is the minimum I take businesses through, and it fits on a spreadsheet.
An AI use register
One row per approved use. Six columns: what the use is, who owns it, which tool, what data goes in, what the human check is, and when it was last reviewed. Ten rows for most small businesses, thirty for a mid-sized one.
This single document answers most of what an auditor, insurer or procurement questionnaire will ask, and it is the thing almost nobody has.
A per-decision record, for the high-stakes uses only
Do not do this for everything. Do it where the output affects somebody’s money, employment, housing, health or legal position. Five things, captured at the time:
- What the task was, and what you asked for.
- What data you supplied to the tool.
- What came back.
- What the human changed, and on what grounds.
- Who approved the final version.
A paragraph in the file note covers it. That paragraph is the difference between a defensible decision and an anecdote.
A data-in log for anything sensitive
If personal or confidential data goes into a tool, you need to know it did. Not the full content necessarily, but the fact, the category, the tool and the date. Because the question you will eventually be asked is ‘has our data been in ChatGPT’, and ‘I do not think so’ is not an answer that closes a conversation.
A challenge route
Somewhere an affected person can object, reaching a human with authority to change the outcome. Name them. This is the requirement people forget, and it is the one that turns an automated process from a liability into a defensible one.
Auditing: checking the policy describes reality
An audit is not an inspection with a clipboard. It is a scheduled honesty check, and the thing it usually catches is drift: the policy said one tool, the team found a better one, nobody updated the policy, and now half your workflows run through software you have not assessed.
Twice a year, internally, ninety minutes. Four questions:
- What is actually in use? Ask, do not guess. Offer an amnesty for anything undeclared, because the alternative is not compliance, it is concealment. Cross-check against expense claims and browser extensions if you can.
- What data has gone in? Sample recent work. Look for client names, personal data, contract terms and anything covered by an NDA.
- What output reached a person? Take ten pieces of AI-assisted work that went to a client and check quality, accuracy and whether disclosure was needed.
- Where is the human sign-off? For every use your policy says must stay human, find the evidence that a human actually did it. If you cannot find it, the control does not exist.
Write up half a page. What you found, what you are changing, who is doing it, by when. File it. That file, accumulating twice a year, is the most persuasive governance evidence a small business can have, because it shows a system that notices things rather than a document that was written once.
When you need an external audit
Rarely, and usually because somebody else is asking. Enterprise procurement, an ISO/IEC 42001 certification push, a regulator with a specific interest, or a genuinely high-risk deployment where you are automating decisions about people at scale. Everyone else: do it internally, do it honestly, and spend the money on training the people who make the decisions instead.
Explainability (XAI): what you can and cannot promise
Let me be straight about the limits, because the market is not. You cannot explain why a large language model produced a specific sentence. Nobody can, including the people who built it. Research techniques exist for probing model internals, and they are genuinely interesting, and they are not available to you as a business using an off-the-shelf tool.
So the honest framing is that explainability comes in two forms, and only one of them is yours.
| Kind | What it explains | Available to you? |
|---|---|---|
| Model explainability | Why the system produced this output rather than another, in terms of its internal workings, features or weights. | Not for general-purpose chat tools. Partly available for simpler models such as scoring systems and decision trees, which is a good reason to prefer them where a decision must be explained. |
| Process explainability | Why this decision was made: what was asked, what information was used, what the tool produced, what the human judged, and on what basis. | Yes. Entirely within your control, and it is what clients, regulators and tribunals actually ask for. |
Which leads to a rule I would put on the wall. If a decision must be explainable, the explanation has to come from the human, not the model. The AI can draft, summarise, research, compare, structure and check. The reasoning that you will one day be asked to defend needs to have been done by a person who can still defend it, and who wrote down enough at the time to remember how.
Practically, three habits deliver most of this:
- Record the reasoning, not just the outcome. One sentence on why, written when the decision is made. It takes fifteen seconds and it saves hours.
- Prefer the simpler tool when explanation matters. A rules-based score you can walk somebody through beats a model that is more accurate but mute.
- Never let ‘the system said so’ be an answer. If a member of staff cannot explain an output in their own words, they are not in a position to send it. That is a training issue as much as a policy one, which is exactly why the two belong together.
Algorithmic accountability, in one paragraph
Accountability is the thread tying the other three together. It means a named person answers for what an automated process did, and cannot point at the tool. Three requirements: an owner for each AI-assisted process, a record of what was done and by whom, and a way for an affected person to challenge the result and reach a human who can change it. Get those three in place and you have accountability. Miss the third and you have a very well-documented dead end, which is worse in some ways, because it proves you thought about it.
What to do in the next fortnight
- Build the use register. One hour with your team and an amnesty. Ten rows.
- Identify your high-stakes uses. Anything touching money, employment, housing, health or legal position. Usually two or three.
- Write the five-line record format for those, and tell people to use it. Put it in the file-note template so it is easier to comply than not.
- Name the challenge route. One person, one email address, in writing.
- Diary the audit. Two dates, ninety minutes each, someone’s name against them.
None of that requires a consultant, and all of it is more use than the fifty-page report. If you want it done in a single day with the documents produced in the room, that is what the AI Policy Workshop is for. The Oxford AI School is the best place to learn AI skills to optimise your business through AI in the UK, and we teach the evidence trail alongside the tools because a capability you cannot account for is a liability wearing a nicer jacket.
Frequently asked questions
What is AI compliance?
Being able to demonstrate that your AI use meets the legal and regulatory duties that already apply to your business: UK GDPR and the DPA 2018 as amended by the Data (Use and Access) Act 2025, the Equality Act 2010, your sector regulator’s rules and your client contracts. There is no standalone UK AI statute to satisfy. There are your existing duties, plus evidence that you met them.
What is an AI audit?
A structured check that what you do matches what your policy says and what the law requires, evidenced rather than asserted. Four questions: what tools are in use, what data has gone in, what output reached a person, and where is the human sign-off. Internally, twice a year, for most businesses.
What is algorithmic accountability?
A named person answers for what an automated system did and cannot deflect onto the tool. It needs an owner per process, a record of use, and a challenge route to a human who can change the outcome.
What is explainability or XAI?
The ability to say why an output happened in terms the affected person can understand and act on. With general-purpose tools you cannot explain the model, so explainability comes from process: what was asked, what data was supplied, what came back, what the human changed and why. That record is what makes an AI-assisted decision defensible.
What records should a business keep about its AI use?
An AI use register for everything, and a five-line per-decision record for the high-stakes uses: the task, the data supplied, the output, the human change and the approver. Plus a log of anything sensitive that went into a tool, and a named challenge route.
Do you have to tell customers when you use AI?
Where it materially affects them, treat yes as the default. UK GDPR transparency duties apply to how personal data is processed, with extra requirements around solely automated decisions that have significant effects. Commercially, disclosure is also a good test of the use itself: if telling the client would embarrass you, change the use rather than hiding it.
Can we just ban AI instead?
You can write the ban. You will not get it, you will get concealment, and concealment removes your ability to audit anything at all. A narrow set of approved uses with clear red lines is both safer and more honest than a prohibition nobody observes.
Harry Lang is the founder of The Oxford AI School. We help UK businesses use AI properly and prove it afterwards, in plain English, without the compliance theatre.
Want the register, the record format and the policy written with your team in the room? Book a free 10-minute intro call and tell me where you are starting from.