Back to blog

AI adoption is now a governance problem, not a productivity story

•7 min read

Most of the AI conversations I have with UK SME leaders start in the same place: someone has seen a tool save time, and the question is how quickly it can be rolled out. That is a reasonable question. It is just not the one a board should be answering first.

Once staff are pasting customer emails into a chatbot, connecting an assistant to the shared drive or letting a tool draft replies to clients, the organisation has made a set of decisions about data, accountability and risk. Usually nobody made those decisions on purpose. They happened by default, one licence and one browser extension at a time. That is the governance problem, and it does not go away because the productivity gains are real.

The productivity case is not the approval case

A productivity case tells you what a tool can do. An approval case tells you what you are agreeing to. They are different documents, and in my experience most SMEs only ever write the first one.

When a board or leadership team signs off on AI use, it is accepting responsibility for questions like these:

  • What data is the tool allowed to see, and where does that data go once it has been processed?
  • Who decided the tool could be used for this purpose, and can that decision be found later?
  • If the output is wrong, who is accountable for having relied on it?
  • If a client, a regulator or the Information Commissioner's Office asks how a decision was reached, can anyone answer?

None of these are exotic. They are the same questions you would ask about a new payroll provider or a new CRM. The difference is that AI tools are cheap, quick to adopt and often arrive through individual sign-ups rather than procurement, so the questions get skipped.

Decision rights: who is allowed to say yes

The first thing I look for in any organisation adopting AI is a clear answer to "who can approve this?" Not a committee with a grand name, just a named person or role with authority to approve a use case, and a short list of things they must check before they do.

In a smaller business this might be the managing director with input from whoever owns IT and data protection. In a regulated firm it will also involve the compliance function, because sector regulators expect firms to be able to explain how their systems and controls work, whether or not AI is involved.

Decision rights matter because, without them, approval drifts to whoever is most enthusiastic. That is how you end up with a sales team feeding prospect lists into a tool nobody has reviewed, or an HR manager using an assistant to summarise grievance notes. Each choice feels small. Together they amount to a data processing estate nobody designed.

A workable starting point is a simple register of approved AI uses: what the tool is, what it is used for, what data it touches, who approved it and when it will be reviewed. It does not need to be elaborate. It needs to exist and be kept up to date.

Data boundaries: what goes in, and what comes out

Under UK GDPR and the Data Protection Act 2018, putting personal data into an AI tool is processing like any other. You need a lawful basis, you need to understand who the processor is and where the data is held, and for higher-risk uses you are likely to need a data protection impact assessment. The ICO has published guidance on AI and data protection, and it is worth reading before you approve anything that touches personal data at scale.

In practice, data boundaries come down to a few plain rules that staff can actually follow:

  • Classify before you connect. Decide which categories of information may never go into an external AI tool: client confidential material, special category data, credentials, unreleased financials.
  • Check the terms, not the marketing. Consumer and business tiers of the same product often handle data very differently, particularly on whether inputs are retained or used for training.
  • Watch the integrations. An assistant connected to email, file storage or a CRM can reach far more than the person using it intended. Permissions granted to the tool should be no broader than the job requires.
  • Treat outputs as drafts. Anything generated for a client, a contract or a decision about a person needs a human who is accountable for checking it.

If you cannot explain your data boundaries in a paragraph, staff will not follow them.

Audit trails: being able to show your working

The question I find SMEs least prepared for is "can you show me what happened?" If an AI-assisted decision is challenged, whether by a client, an employee or a regulator, you need to be able to reconstruct it.

That does not mean logging every prompt forever. It means proportionate records for the uses that carry risk:

  • Which tool and model version was used, and under which approval.
  • What data was provided to it, at least by category.
  • What the output was, and who reviewed or acted on it.
  • Any exceptions, incidents or complaints linked to the use.

Where AI contributes to significant decisions about individuals, such as recruitment screening, credit or eligibility checks, or performance management, UK GDPR's rules on automated decision-making come into play. These were revised by the Data (Use and Access) Act 2025, which replaced the old Article 22 regime with a broader "significant decisions" model, but the safeguards still apply: you need to tell people, let them contest the outcome and offer meaningful human review. The ICO's guidance on AI and data protection covers the current position, and it is worth checking the latest version before you rely on it. An audit trail is how you demonstrate that the human review actually happened rather than being a box someone ticked.

Many business-grade AI platforms now offer admin logging and retention controls. They are often switched off by default or sit on a higher licence tier. Check before you assume you have them.

What I would ask a board to approve

When I help a leadership team put AI governance in place, I try to keep the first version short enough that people will use it. Typically that means the board approves:

  1. An acceptable use statement in plain English, covering which tools are permitted, what data must never be used, and the expectation that outputs are checked.
  2. A named owner for AI risk, with authority to approve or refuse new use cases.
  3. A register of approved uses, reviewed at a set interval, with a simple route for staff to request something new.
  4. Minimum technical controls: business accounts rather than personal ones, single sign-on where available, logging switched on, and integrations scoped to what is needed.
  5. A review trigger, so that any incident, complaint or significant change in a tool's terms brings the use case back for reconsideration.

None of this slows adoption down in any meaningful way. What it does is turn a collection of individual experiments into something the organisation can stand behind when it is asked difficult questions.

Productivity still matters, but it is not the control

I am not arguing against AI adoption. Used well, these tools genuinely help small teams do more with less, and SMEs that ignore them will feel it. My point is narrower: the benefits are not a substitute for governance, and a board that approves AI on the strength of a time-saving demo has not really approved anything.

The organisations that get this right treat AI the way they would treat any other system that handles their data and influences their decisions. They decide who can say yes, they draw clear lines around data, and they keep enough of a record to explain themselves later.

If you are a UK SME leader working out how to approve AI use without either blocking it or losing control of it, get in touch through my contact page. You can also read more about my services and security strategy work, including how I help leadership teams put proportionate AI governance in place.