“You’re Absolutely Right!” Now Who Is Governing Your AI?

AI rarely arrives by decision. In most organisations it crept in. Someone tidied up a difficult letter with a chatbot, and a supplier switched on an assistant inside software you already pay for. The email announcing it is probably still unread.

Deloitte asked around 25,000 UK workers about this in September. Half of those using AI at work had been given no formal guidance or training. Nearly one in ten of everyone surveyed had used a tool their employer had banned or would frown on. The National Cyber Security Centre's annual review, published this month, warns that this kind of unapproved use leaves an organisation open to attack, because nobody can see where its information has gone. So when a client questionnaire asks how you govern AI, the honest answer in a lot of places is that nobody does, yet.

What happens when nobody is checking

The clearest examples this year have come from the legal profession, because judges read the footnotes. In May, a judge in Dudley found that appeal documents prepared by a small firm cited cases that did not say what was claimed. The judge concluded they had been produced using AI. A paralegal had helped prepare them, and the supervising solicitor admitted he had not checked the research. The judge decided a public telling-off was not enough and referred him and the firm's director to the SRA. If anyone had pushed back on the tool along the way, it would no doubt have apologised and told them they were absolutely right.

This year the Solicitors Disciplinary Tribunal also struck off a foreign lawyer who had used AI to draft his own defence. When the SRA pointed out that his citations were wrong, the replacements he offered were little better, and his email explaining the mistakes, also drafted with AI, contained more of them. Nobody suggested he meant to deceive anyone. What concerned the tribunal was that he kept failing to check.

What the two cases share is a process with no point at which a person confirms the output before it counts. Swap the courtroom for a board paper and most organisations have the same gap.

The law got there first

It is easy to assume there are no rules because the UK has not passed an AI law. There are plenty. They were simply written for something else. A dedicated law may yet come: one parliamentary committee has called for an AI bill, and another has opened an inquiry into who is responsible when AI causes harm. For now, the existing rules are the ones that apply.

If a tool handles personal data, data protection law applies to it exactly as it would to any other system. Where a tool makes decisions about people, such as which job applications go forward, the rules changed in February this year. Those decisions are now allowed more widely, but the person affected has to be told, and has the right to have a human look at it again. The ICO (formally the Information Commission since September) must now produce a statutory code on AI and automated decisions, expected next year, and in October it began consulting on AI agents, the tools that take actions as well as answer questions. Its draft guidance makes a point worth repeating: buying the tool from someone else does not shift the responsibility to them.

Equality law applies too. If a screening tool marks down applicants with career breaks, the Equality Act 2010 does not care that a machine made the call. The organisation using it answers for the outcome.

If you are regulated, your regulator has already said what it thinks. The FCA decided last year not to write AI rules and to hold firms to the ones it has, with a named senior manager accountable for how AI is used. The SRA's warning notice to law firms in August put it more plainly: if you used AI to do the work, it is still your work.

Across the Channel it is different. The EU has an AI Act, and it reaches UK organisations whose AI is sold into the EU or whose output is used there. Its heavier requirements, for uses such as recruitment and credit, now start in December 2027 after a delay agreed this summer. If that might apply to you, talk to a lawyer rather than guess.

Clients are asking anyway

In practice the pressure usually comes from clients before regulators. We have watched it arrive in our own client meetings. A year ago, having an AI policy answered the question. This autumn two clients raised it with us, unprompted, in the same fortnight. Supplier questionnaires that used to stop at cyber security now ask which AI tools you use and whether client information goes into them. Some contracts forbid AI on client work without consent. None of that is law, but once you have signed it binds you just the same, and a vague answer reads badly.

What governing it actually means

It sounds grander than it is. At heart, it means you can say without a week of emails what AI is in use, including the features that turned up inside existing software, and that someone has decided what is allowed. The part people forget is the checking. Where AI shapes a decision about a person, somebody has to look at the output before it counts. The NCSC put much the same point to boards this month: as AI tools start to act for themselves, an organisation needs to understand how they are being used and what they can reach.

In our experience the list of AI in use is the piece most organisations underestimate, because the AI that matters most is often a feature that appeared inside software they already had, which nobody chose. A single line in that register might read:

Alongside it, a short policy. Two pages is plenty if it is clear about which tools are approved and what must never go into them. It also pays to agree one standard set of answers to the AI questions clients ask, changed only with sign-off, so every bid says the same thing. None of this needs new software.

There is a standard for this, which surprises people

Plenty of people do not know there is an international standard for governing AI. ISO/IEC 42001 was published in 2023 and takes the basics above much further, into how you judge the effect of AI on people and keep a person in control. Fewer still have heard of ISO/IEC 27090, published in October 2026. It belongs to the ISO 27001 family and deals with security: someone poisoning what a tool learns from, for example, or instructions hidden in a document that the tool then quietly follows.

The reassuring news is that little of this is new work. When we mapped 27090 against the ISO 27001 controls, most of the 28 areas it covers were already handled, fully or in part, by controls organisations have today. Only three were real gaps, mostly to do with spotting and withstanding deliberately manipulated inputs.

42001 is similar. For an organisation already running ISO 27001 it extends what is there, because the two share most of their machinery. What is genuinely new is an assessment of the effect on people (ISO/IEC 42005 sets out a method) and 38 controls of its own, chosen through a separate Statement of Applicability.

The easiest way to tell them apart: 27090 asks whether your AI can be broken into, and 42001 asks whether you should be using it that way at all. Here is how they sit next to the one piece of actual AI law in our region.

Where most organisations should start

For most, the first step is unglamorous. Find out what is actually in use and make someone responsible for it. If you hold ISO 27001, put the AI threats from 27090 into your risk assessment and the resulting controls into your Statement of Applicability, and most of the work already has a home. ISO 42001 tends to earn its place later, once AI starts shaping decisions about people or clients ask for it by name. Nothing done before then is wasted.

An auditor who finds AI in use will usually start by asking whether you are using it or providing it to others. That answer can change quickly. Once you deliver AI-enabled services to clients you become a provider, and both 42001 and the EU AI Act expect more of you.

Here is a quick test of whether any of it is working. Pick an AI tool at random and ask who approved it and when someone last checked what it produced. Then, next time a questionnaire asks how you govern AI, notice whether the answer takes ten minutes or ten days.

Frequently asked questions

Not sure where you stand?

Thirty minutes with a practitioner will usually settle it, with no proposal at the end of the call. If you would rather start on your own, the free Readiness Check gives you an indicative SSF Effectiveness Score in minutes, with no documents required.