The AI seat is not the mistake. Buying it first is.

The Department for Business and Trade gave out 1,000 Microsoft 365 Copilot licences for a three-month pilot, and published the results. Microsoft's own dashboards showed about 30% of licence holders using it on a given working day, and about 64% at least once a week. On a typical working day, about seven in ten seats went unused. The evaluation collected little data on output, and found no evidence that the time saved turned into any. For a 50-350 person company, buying AI seats per head before you know what work they are for gets the order wrong. A seat helps one person write faster. The money is in specific work between your systems, and a seat will not find it for you.

What a seat actually buys

Some of the best public evidence on seats comes from the UK government, which has trialled Copilot at scale and published what it found. The cross-government experiment, reported in June 2025, covered 20,000 staff across 12 organisations. Participants reported saving 26 minutes a day, and 82% said they would not want to go back. The DWP's own evaluation found 19 minutes a day across eight routine tasks. Both figures are self-reported, and the DWP trial was not randomised.

Even in the Department for Business and Trade pilot, where most seats went unused on a given day, 72% of respondents were satisfied.

People like it. That is not the same as it paying.

What the seat costs is easier to pin down than what it returns. At list price, the Copilot enterprise add-on is £23.10 per user per month on an annual plan, excluding VAT. For 200 staff that is £55,440 a year, paid whether the minutes saved turn into anything or not.

Why seats go quiet

A seat with no named job gets used for whatever comes to mind, and then less. In a 2024 Gartner survey of 132 IT leaders, reported by Computerworld, 72% said employees struggled to fit Copilot into their daily routines, and 57% saw engagement drop off quickly. Only 1% had rolled it out to every eligible office worker.

Meanwhile, staff are finding AI on their own. MIT's NANDA research found workers at over 90% of companies using personal chatbot accounts, while only 40% of companies had official subscriptions. Microsoft's own Work Trend Index found in 2024 that 78% of AI users were bringing their own tools to work, rising to 80% at small and medium companies.

So the general-purpose assistant is often in the building already, on staff's personal accounts. A company licence brings it inside your data controls, which matters. It does not, on its own, give anyone a different job to do. The ONS finds that UK firms using AI have adopted an average of around 1.6 AI technologies, which, it says, implies "relatively limited transformative impacts to date".

Buying seats changes the invoice, not the work.

Where the money is

The firms getting a return are the ones that changed the work. BCG's Applied AI Index, published on 30 September, found that almost half of companies now generate value from AI, and that the leaders "apply it in high-value areas by reshaping workflows end to end". MIT's study, as reported by Fortune, found the biggest return in back-office automation, even though more than half of budgets went on sales and marketing tools.

In a company of 150 people, that work rarely looks like writing. It is an order that arrives by email and gets retyped into the ERP. It is a supplier invoice checked by hand against a purchase order. It is a weekly report built by exporting from three systems into one spreadsheet. None of it lives in one person's day. It is spread across roles and handed between tools, so no single seat ever sees it from end to end.

The return lives between your systems, not in anyone's inbox.

That work has to be found before it can be fixed, and then costed: how many hours a week, whose hours, how many errors, how much delay. Only then do you know what fixing it is worth, and whether it is worth more than a licence. No seat will do that counting for you, because no seat can see the whole of the work.

When the seat is the right buy

None of this makes seats a bad buy. For people whose work is mostly drafting, reading and summarising, a general assistant is cheap and does what it says. Named jobs help: BBVA, which tied its rollout to specific uses, says 83% of its licensed staff use ChatGPT daily and that over 3,000 task-specific assistants have been built. Named jobs make seats useful; they do not reach the work between systems.

Seats are also getting cheaper to try. Gemini now comes included in Google Workspace Business Standard, so some companies already have a seat they have not noticed. Claude Team is priced in dollars, at $20 per seat per month on annual billing, and is capped at 150 seats, which suits a department rather than the whole of a larger company. Even Microsoft is moving away from pricing purely by head: in July, Satya Nadella spoke of evolving "beyond per-seat to per-seat-plus-consumption".

Is it worth doing? Not first. Find where people do manual, repetitive work between systems, and put a number on what it costs. Fix the largest. Then buy seats for the people whose work is mainly general writing, where a seat is cheap and sensible.

Buy the seat. Just buy it second.