‘We have Copilot’ is not an AI strategy

Somewhere this week, a board is being shown an AI strategy that is one slide long. The slide says Copilot.

Nobody on that board is being lied to, exactly. Licences were bought. The rollout happened. People are using it. Every word of the slide is true, which is precisely what makes it so hard to challenge. The problem is what the slide is standing in for.

The numbers say adoption. They do not say depth

The Office for National Statistics published its analysis of AI in UK businesses on 20 July 2026, drawing on the Business Insights and Conditions Survey with fieldwork run 15 to 28 June 2026 and 38,637 businesses responding. The headline is genuinely impressive: the share of UK businesses with ten or more employees using at least one AI technology has almost tripled since late 2023, from around 12% to around 35%.

Then comes the sentence underneath the headline. Over the same period, the average number of AI technologies used per adopting business rose from around 1.4 to around 1.6. Adoption nearly tripled. Depth moved by a fifth of a technology.

The ONS asked a new question this wave that makes the point sharper still. Among businesses using AI at all, only 10% describe their use as extensive. The rest are, by their own account, dabbling.

The CIPD’s analysis of its Autumn 2025 Labour Market Outlook, published in January 2026 with fieldwork run in autumn 2025, shows the same pattern from the employer’s side. In the previous twelve months, 43% of employers had introduced staff platforms that use generative AI, naming Copilot specifically as the example. The share who had identified activities or business use cases where generative AI could be applied was 36%. Sit with that ordering for a moment. More organisations deployed the tool than worked out what it was for.

Meanwhile your staff got on with it

The same ONS release compares what businesses report with what individuals report. In fieldwork run May to June 2026, 55% of employed people said they use AI for work or education. Businesses reporting use of at least one AI technology: 35%.

That gap is your shadow AI estate. It is your analyst pasting customer data into a free chatbot because the sanctioned tool cannot do what she needs. It is the report half-written by a model nobody evaluated, on terms nobody read. The individual adoption curve is ahead of the organisational one, and where those curves diverge, the difference is made up of usage nobody is accountable for.

A licence rollout does not close that gap. It gives the shadow usage a nicer logo.

Why the licence feels like a strategy

I have some sympathy for the one-slide board pack, because buying licences is the only part of AI adoption that behaves like a normal IT decision. There is a vendor, a price per seat and a procurement process that knows what to do with both. It produces an artefact you can point at. Done.

Deciding what AI should actually do in your business behaves nothing like that. It requires knowing your own processes well enough to name the ones that are broken, which is uncomfortable. It requires saying no to most of the candidate ideas, which feels like slowness. It produces its value later, which is a hard sell against a rollout that produces a completion date now.

The ONS has previously analysed what holds adoption back, and the most commonly reported factors preventing or delaying AI adoption included difficulty identifying business use cases, alongside cost and lack of expertise. Read that against the CIPD finding above. The hard part is not access to the technology. The hard part is the decision, and the decision is exactly what a licence purchase lets you skip.

What an adoption decision actually looks like

We run a method for this. We call it APEX, six phases: Identify, Discover, Analyse, Design, Implement, Monitor, with a quality gate at the end of each. It exists because we kept watching the same failure, which is organisations starting at phase five. Buy the tool, implement it, then wonder why nothing measurable changed. The declared interest is obvious, we sell this. The shape of the process is worth having whether you buy it from us or build your own.

Identify is a shortlist, not a brainstorm. Everyone’s first AI workshop generates forty ideas. The useful output is the thirty-five you strike off. We score candidate processes against five weighted criteria, from business impact through to stakeholder readiness, before anything gets built, so the choice of where AI goes is a decision with reasons attached, not a reflection of who spoke loudest in the workshop.

Discover means mapping the process as it actually runs. Not the version in the procedures folder. The version with the workaround Karen invented in 2021, the approval that waits for someone’s Tuesday, the spreadsheet nobody else can open. We model the as-is process properly, in BPMN, because you cannot safely automate a process you have not honestly described. This phase is where most of the surprises live. Better here than in production.

Analyse and Design decide how much agency the AI gets. The working principle is minimum viable autonomy: the smallest amount of independent action that still delivers the benefit. Drafting a reply is a different risk from sending one. Recommending a decision is a different risk from making it. Most of the value in most processes sits at the safe end of that scale, which is convenient, because the safe end is also where legal and security will actually sign.

Implement and Monitor treat go-live as a beginning. The system that watches the AI is part of the build, not an afterthought. If nobody owns the question “is it still behaving?”, the honest answer eventually becomes no, quietly.

None of this is exotic. It is the same discipline you would apply to any operational change that touches customers and money. The only unusual thing is applying it to AI, where the industry default is to let enthusiasm skip the queue.

What to do with the slide

If your AI strategy currently fits on the Copilot slide, some free suggestions, in order.

Ask what the licences are for. Not rhetorically. Ask which specific processes they are meant to improve and what would show it. If the answer is “general productivity”, that is the sound of a decision not yet made. The government’s own trial of general-purpose AI assistance found civil servants saving an average of 26 minutes a day, which is real and worth having. Notice that measuring it required deciding what to measure.

Find out what your people already use. The gap between the 55% and the 35% exists inside your organisation too. The staff who adopted AI before you did are telling you where the friction is. Their workarounds are a free discovery exercise.

Pick one process. Not a function, not a department. One process with a beginning, an end and a cost you can name. Score it honestly: does it happen often enough to matter, is the data behind it fit to automate, what happens when the AI gets it wrong. Then decide how much autonomy the smallest useful version needs.

Only then talk about tools. The tool question is real, but it is the last question, and by the time you reach it honestly it mostly answers itself.

The gap in the UK numbers is not an adoption gap. Thirty-five per cent of businesses have adopted something. It is a decision gap, between organisations that bought AI and organisations that chose what it is for. The second group is going to be difficult to compete with.

—

Occasional, useful notes on applied AI: what’s actually working, what to ignore and what the new regulation means for UK businesses. Subscribe to Insights via the form in the footer of aiapplied.uk. No spam.