A pattern I see in almost every business I walk into: a company-wide AI licence rolled out a year or two ago, a handful of enthusiasts using it heavily, most people using it to draft emails, and the core processes (quoting, onboarding, reconciliation, reporting) exactly as they were. If you're looking at a renewal invoice and suspect your AI licences are not delivering value, you're probably right. More training or a new vendor won't fix it. What fixes it is deciding which work should change, and building that change into the systems where the work happens.
Licences don't change workflows. Someone sitting with the person doing the work does. Workflows change when someone maps where the hours go, builds the thing that removes them, and stays until the new way is the habit. That isn't a training problem and it isn't a strategy problem. It's a build and embed problem.
Licences don't change workflows. Someone sitting with the person doing the work does.
Why Copilot adoption stalls after the launch email
The national numbers say the same thing as the renewal invoice. The ONS reported in July 2026 that self-reported AI use among UK businesses with ten or more employees rose from around 12% in late 2023 to around 35% in June 2026. That's roughly tripled. But the average adopting business uses only about 1.6 AI technologies, up from about 1.4, and large language models are the most used, by 18% of businesses. In plain terms: a lot of firms bought a chat assistant, and not much else happened.
I don't think that's because people are lazy or because the tools are poor. Microsoft Copilot, ChatGPT Enterprise, Gemini and the rest are impressive. The problem is structural, and it shows up in four places.
The chat window sits beside the work, not inside it. A credit controller's day lives in the ledger, the inbox and a spreadsheet of disputed invoices. The assistant can help her write a firmer chasing email. It can't open the aged debt report, match remittances, flag the three accounts that always pay late and draft the chase before she arrives. That needs something wired into the systems, with rules the business has agreed.
Nobody redesigned a process. The rollout was an IT project with a procurement decision at the front and a launch email at the back. Nobody sat down and said "quoting will now work like this". So quoting didn't change.
Training taught prompting, not the new job. Most rollout training I've seen is a lunchtime session on writing better prompts. Useful, in the way a typing course is useful. It doesn't show an estimator how their week should look when the first draft of every quote arrives pre-filled from the job sheet. Training that works has people build something for their own job, which I've set out in how to train staff on AI.
No owner and no baseline. Nobody was accountable for hours returned, and nobody measured the process before the licences went live. So when the CFO asks what the spend achieved, the honest answer is "people like it". MIT NANDA's 2025 study found 95% of the generative AI pilots it looked at showed no measurable P&L impact, and it put the core cause down to a "learning gap": tools that don't retain feedback or adapt to the workflow. Critics have pointed out that part of "no measurable impact" is simply that no baseline was taken. Both readings fit what I see in licence rollouts.
A licence audit a COO can run in a week
Before anyone decides whether the rollout was a waste, get the facts. This takes a week of part-time effort and nobody external.
Day one, pull the usage reports. Every major assistant has an admin console that shows active users, frequency and roughly what features they use. Export 90 days, not the last week, so holidays and month-end don't distort it.
Day two, segment the seats. I use three buckets, because the decisions are different for each.
| Segment | What it looks like in the report | Seats in a 150-seat example | What to do |
|---|---|---|---|
| Daily power users | Active most working days, long sessions, several features | 15 | Keep. Interview them this week. |
| Occasional users | Active a few times a month, mostly drafting and summarising | 35 | Keep for now. Review at renewal. |
| Dormant | Logged in once or twice at launch, little or nothing since | 100 | Reclaim at the next renewal point. |
Those figures are illustrative, but the shape is common. A small core uses it constantly, a middle group dabbles, and most seats are paid for and idle.
Days three and four, interview the power users. This is the most valuable part of the week and most businesses skip it. Sit with five or six of them for thirty minutes each and ask them to show you, on screen, what they did with it yesterday. Not what they think it could do. What they did.
You're listening for repetition. The operations coordinator who pastes the same supplier emails in every morning to pull out delivery dates. The account manager who turns every call transcript into a CRM update by hand. The finance analyst who uses it to reconcile two exports that never quite match. Each of those is a person who has already found a process worth automating, and has been quietly doing it with a chat window and copy and paste. They are your best process candidates, and they've done the discovery for you.
Day five, do the cost arithmetic and decide. Take the illustrative example. Say you pay £25 per seat per month (use your own contract figure, which will differ). 150 seats is £3,750 a month, or £45,000 a year.
Divide that by the people who genuinely use it. With 50 active users (the 15 daily plus the 35 occasional), the effective cost is £45,000 ÷ 50 = £900 per active user per year, or £75 a month. That's three times the list price you thought you were paying. If you're stricter and only count the 15 daily users, it's £3,000 per user per year.
Now reclaim the 100 dormant seats. The same 50 people cost you £15,000 a year. The business loses almost nothing, because those 100 people weren't using it. You've found £30,000 a year before you've built anything, and you have a list of the processes your keenest people are already propping up by hand.
Check your contract terms before you promise the board a saving. Some agreements commit you to a seat count until renewal. The audit still tells you what to do at that point.
From assistant to workflow
The audit usually surfaces five to ten candidate processes. Pick one or two. I'd choose on volume (how often it happens), time (how many minutes each instance takes) and pain (how much rework, delay or risk it causes). The one that scores highest on all three tends to be dull: supplier confirmations, job sheet to invoice, onboarding paperwork, the weekly pack.
Then build it properly. That means inside the systems the work already lives in, not another tab. The supplier email arrives, the delivery date is extracted and written into the ERP, exceptions go to a named person with the original attached, and the coordinator only sees what needs a human. The assistant might still be part of it (some of these builds use the same underlying model you're already licensed for), but it's now doing a defined job in a defined place with a defined owner.
The practical answer is a short embedded sprint: two to three weeks inside the business, the top one or two processes automated and running before the end, the team watching it happen so they own it after. The power user who found the process sits with whoever builds it. They know the edge cases. They know which supplier sends dates in American format and which customer always changes the order after confirmation. That knowledge is the difference between a demo and something that runs.
A few rules I hold to on these sprints.
- Measure the process before you touch it: volume, minutes per instance, error rate. Without that you're back to "people like it".
- Build for the exceptions first. The easy 80% is the demo. The awkward 20% is what decides whether people trust it.
- Hand it over to a named internal owner before the sprint ends, with a one-page note on what it does and what to do when it breaks.
- Check data protection early. If personal data moves between systems, your UK GDPR position and your ICO obligations don't change because a model is involved.
It costs a fraction of a transformation programme and it produces something you can measure in the following month's numbers. If you want the fuller sequence for building several of these over a quarter, I've set it out in a 90-day plan for implementing AI in a mid-market business. This is the smaller, sharper version: start from what your licence holders already found.
What to tell the board about Microsoft Copilot ROI
At some point someone on the board will ask what the AI spend achieved. Don't answer with log-in charts. Seat activation is an activity measure, and a board that has seen a few technology programmes will know it.
A better update fits on half a page. What you paid, what the effective cost per active user turned out to be, how many seats you're reclaiming and when. What the power users are doing with it, in one or two concrete sentences. Which one or two processes you're now building into the real systems, the baseline you measured, and the month in which you'll report hours returned to the business. That last number is the one that matters, and I've written about how to make it survive a sceptical CFO in measuring AI ROI.
Be straight about the licence spend itself. A general assistant is a reasonable thing for a knowledge business to pay for, in the same way a decent laptop is. It won't show up as a line in the value creation plan, and nobody should pretend it will.
When licences are the right tool, and when they're a substitute for a decision
General assistants are genuinely good at some things. Drafting a first version of a letter, a proposal or a policy. Researching a topic quickly and summarising a long document. Taking meeting notes and producing actions. For people whose job is mostly writing, reading and thinking, a seat can easily pay for itself in a month, and I'd keep it without much argument.
Where they go wrong is when a company-wide rollout stands in for a decision nobody made. "We've given everyone Copilot" can be a way of saying "we've done AI" without choosing which process should change, who owns it, or what good looks like in a year. It feels like action. It's a purchase.
The test I use is simple. If the benefit depends on each individual remembering to open a chat window and paste something into it, it's a personal productivity tool, and you should judge it as one. If the benefit depends on work flowing differently between people and systems, it needs building, and a licence on its own won't do it.
ChatGPT Enterprise adoption and Copilot adoption figures look good on a slide precisely because they measure the first kind of benefit. The second kind is where the hours are.
The six-month test
S&P Global Market Intelligence found that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. Much of what gets abandoned was never wired into anything, so it was easy to switch off. A licence rollout is the easiest of all.
So the measure of success is boring on purpose: what is still running, unattended, six months after the outside help has left. Not how many people logged in during launch week. Not how many prompts were written. Which processes now run differently, without someone nursing them, and how many hours a month they return.
If you run the audit and find that the answer today is "none", that's fine. It's a clear starting point, and your power users have probably already told you where to begin. If you'd like help turning what they found into something that runs, that's what embedded AI delivery is for.
Questions
Why is our Microsoft Copilot rollout not delivering value?
Usually because the assistant sits beside the work rather than inside it. People use it to draft emails and summarise documents, which helps them individually, but nobody redesigned a process around it, nobody owns the outcome and nobody measured the baseline. Core processes such as quoting, onboarding and month-end stay exactly as they were, so the P&L never sees a difference.
How do I measure Copilot or ChatGPT Enterprise adoption properly?
Pull the admin usage reports and split users into daily power users, occasional users and dormant seats. Divide the total annual licence cost by the number of genuinely active users to get the effective cost per active user. Then interview the power users about what they actually use it for. Log-ins are an activity measure; changed processes and hours returned are the ones that matter.
Should we cancel our AI licences if usage is low?
Not all of them. Keep seats with the people who use the assistant every day and reclaim dormant ones at the next renewal. General assistants are genuinely good for drafting, research and meeting notes. The mistake is treating a company-wide licence as the whole AI plan. Use the audit to find one or two processes worth building properly, then decide the seat count.
What should I tell the board about our AI licence spend?
Tell them the effective cost per active user, what the heavy users do with it, how many seats you are reclaiming, and which one or two processes you are now building into the real systems. Give a baseline and a date by which hours returned will be reported. Boards cope well with honest numbers and a plan. They cope badly with seat counts presented as progress.
