.png)
An advisor I work with told me a story this week. His client's consultants had been asking for Claude for months. "We need it, we need it, we need it." Eventually the client said yes, on one condition: each consultant had to write down what they would use it for.
"Crickets."
That one word is the whole story of AI adoption in the mid-market this year. Months of asking, and not one person could spend an hour describing how they would actually use the tool. Almost nobody is telling this story, because everyone in the supply chain — vendors, resellers, advisors — is paid when the licence is sold, not when it is used.
Here is the idea this article rests on. A request costs the person making it nothing. A use case — a plain description of how the tool will change their own job — costs them an hour of genuine thought. Organisations are sizing their AI estates against the first number, and it is budget season, so they are doing it right now.
Four conversations in five days
The consultants were the first. I heard the same story three more times that week, in three organisations that have nothing to do with each other.
The head of cloud and workplace at a 1,500-seat water utility described his own licence estate with more honesty than most vendors manage in a year: "We've done really bad design around going, last person to leave, next person on the list now then gets the co-pilot licence without actually looking at the use case ... So people aren't using it." He is now writing a business case to lift from 225 licences to 400. And in the same conversation, he named the reason that business case should worry him: "We've got people that want it, but they don't know why they want it."
An IT director at a 4,000-staff aged care and community services provider had already done the arithmetic that rules out the easy answer: "If we had to buy a co-pilot licence for everyone, we'd be up for 1,000,000 bucks." He was never going to do that. So his real question is who should get a licence, and he was candid about how hard that is to answer: "I'm interested as to if I can get other people interested in it. That's going to be a challenge in amongst itself." He is not worried about the technology. He is worried about finding people in his own organisation who genuinely want to use it. Right now, the most interested person in the building is him.
The third did something different. The head of IT at a large document technology business, running a heavy SAP estate, refused to take the ask at face value: "Last year we did a bit of like investigation and discovery around what can we use Copilot for ... Let's not just buy and give it to everyone." Today he has "100 and some people using it." Note the word. Not licensed. Using. When staff asked about consumption-based agent billing, he held the line: "I've had a couple of people ask me about it and I'm like, no." A couple of people asked. He did not build a programme around them.
What organisations are getting wrong
The first two organisations are making the same two mistakes. They are treating the request as demand, and the licence count as adoption.
That number is not demand. It is interest. And interest disappears the moment someone asks a question that costs anything to answer. The hour it takes to write a use case is the most expensive hour in any organisation, and it is where every stalled AI programme I have seen this year quietly died.
The same gap opens again one step later. Having an idea costs nothing. Turning it into a working process costs someone ownership, governance, and their name on it when it breaks. So there are two free signals — requests and ideas — and organisations are sizing their AI estates against both.
The research shows what happens next. KPMG's 2026 pulse survey found about 60% of organisations evaluate a custom AI tool, 20% pilot one, and around 5% put one into production. MIT's much-quoted finding that 95% of pilots show no P&L impact lands at the same one in twenty. Those are not numbers about a shortage of ideas. They are a conversion rate.
The industry's standard answer is enablement: more training, more champions, better change management. I have sold that answer myself. It works when demand is real but blocked. It does nothing when the demand was never there, and you only find out which one you had after you have bought the seats.
Where real demand hides
There is a second signal in these same conversations, and it points the other way.
The IT director at the aged care provider mentioned, almost in passing, that he had built himself a small agent to answer policy questions. "I created this little thing called the Policy Bot ... I don't share that with anybody yet."
That is what real demand looks like in a mid-market business: private, specific, unadvertised, and invisible to every survey you will ever run. Somebody wanted the outcome badly enough to give up a weekend for it. He was looking for people who genuinely wanted the tool, and the one person who did was him.
Meanwhile the loud demand — hands in the air, the email to IT, the line in the engagement survey — is what your licence model is sized against. You are buying the signal that costs nothing to send and ignoring the one somebody paid for.
If you want one number to run your AI programme against, it is not seats and it is not survey scores. It is the number of people in your business who have built something and not told you about it.
And if you are a provider, as I am
This part is uncomfortable, so I will start with my own number.
We run an internal AI skills programme to build capability across our team. This month it stalled. The reason people gave was ideas. "A few people are starting to run out of ideas," one colleague said. Another agreed within minutes: "Running dry on some ideas."
I believed that for about a day. Then I realised it was not true. Nobody here is short of ideas. What we are short of is the step after the idea: turning something that worked once in a chat window into a process that is owned, governed, measured, and still running in ninety days.
So a company that sells AI enablement had been counting the free signal too. Our clients count requests. We count ideas. Same error, one step apart.
If this happens to us, it will happen to a 300-person logistics firm whose provider has just sold them a champions programme. The honest thing a provider can do is stop counting hands and start counting processes that are still running ninety days later, and agree to be paid on use rather than on seats. Almost nobody in my industry will say that out loud, because the request is where the margin is.
Three questions before you sign the licence uplift
- What happened the last time you said yes? Not what was requested — what was built, used, and still in use ninety days later. If you cannot answer that, the requests in front of you are not evidence of anything.
- What is your licence allocation actually based on? If a seat passes from the person who left to the person who joined, you are not allocating against value; you are maintaining a headcount ratio. Make a named use case the unit of allocation, and watch how many seats you actually need.
- Who has already built something and not told you? Find them before you commission a training programme. One person quietly running their own agent is worth more than fifty people saying yes in a workshop.
Demand that costs nothing to express is not demand. It is politeness.
Price the signal. The request, the idea, the hand in the air — none of it counts until someone has paid for it with an hour of thinking about their own job. What survives that test is your real demand. It is a smaller number than the one in your business case, and it is the only one worth buying against.