A client of ours made three separate AI purchasing decisions last week. We were involved in one of them, and only as the order desk. We are still their managed service provider.
They had engaged a SharePoint consultancy. They had a training company coming in to run Copilot education. And they had bought two Agent 365 licences through us — a clean transaction, processed without anyone asking whether two agent licences were the right shape of the answer, or whether an E7 step-up would have served them better. Three separate decisions in a single fortnight, and the first I knew of the other two was a passing mention in conversation. Nobody was hiding anything. It simply had not occurred to them that their MSP was part of that category of purchase — and on the one purchase that did come through us, it had not occurred to us to ask the bigger question.
That is the thing worth sitting with, and it is not really a story about us. It is what disintermediation actually looks like in the mid-market. It does not arrive as a tender or a termination notice. It arrives as small invoices you were never copied on — and order-taking on the one you were.
The shape of the problem
For twenty years an MSP's protection was structural. Infrastructure purchases were large, slow and visible. You could not migrate a mail platform or replace a firewall without the provider in the room, because the provider held the keys. That gatekeeping was never a strategy, but it worked like one.
AI purchases in a small or mid-sized organisation break every one of those conditions. They are small enough to sit under an approval threshold. They are fast enough to be done in an afternoon. They are bought by a function — marketing, finance, a data team — rather than by IT. And critically, they do not feel like infrastructure. Nobody rings their MSP to ask permission before booking a training course.
So the provider keeps the contract, keeps the tickets, keeps the monthly invoice, and quietly stops being the organisation's technology adviser. The relationship is not lost in a review. It is lost in the gaps between reviews.
Three signals from a single week
The first is the one above: three AI decisions made by a managed-service client without a strategy conversation — two with vendors we never saw, one transacted through us without anyone asking the wider question. Not out of dissatisfaction. Out of category confusion — they do not think of their MSP when they think about AI.
The second came from a hundred-person engineering software business I spoke to the same week. Capable internal IT team, currently recruiting more. Around twenty self-appointed AI champions across different departments, each running their own use cases on their own tools. No data governance work started, no AI strategy yet, and — his words — "very minimal" visibility of what any of them are doing. When I asked what triggered the executive conversation about an AI strategy, the answer was not a regulator and not a breach. It was that their own clients had started asking how they handle data and governance.
The third was the most instructive, because it was said out loud. An executive at an 800-person health network, describing the incumbent provider who supports their Microsoft environment and has been helping with Copilot education: their knowledge "is higher than ours, but not where it could take us into the future. And I think that's where to buy that." That is not a complaint. It is a procurement decision being made in real time, in a sentence, about a supplier who will never see the meeting where it happens.
Why the usual diagnosis is wrong
The industry has a word for this and the word is wrong. We call it shadow AI, and the shadow-AI literature is written for organisations without a mature IT function — the implied fix is more control, more policy, more visibility.
But look at where it is actually happening. It happens fastest in organisations that have capable people, because capable people adopt capable tools first. An AI champions programme does not contain the sprawl; it sponsors it. You have given twenty people a mandate to experiment and no framework for what happens next, and they will each pick a different tool, a different budget line and a different vendor. Competence accelerates the fragmentation. That is the uncomfortable version, and it is the one nobody selling AI governance wants to lead with, because it implies the problem is not incompetence to be fixed but momentum to be steered.
What this means if you are the buyer
Do a two-hour audit this week. Ask finance for every payment in the last six months to anything that could plausibly be AI — a licence, a training course, a consultancy, a pilot, a subscription with the word "assistant" in it. Then ask a single question of the list: who made sure these decisions fit together?
That question matters more than the total, because small AI purchases are not small architectural decisions. A training company decides what your staff believe is possible. A SharePoint consultancy decides how your content is structured, which decides what an AI can retrieve later. An Agent 365 licence decides how agents get identities and who audits them. Three decisions made in isolation have just set your identity model, your data surface and your cost model, and nobody is accountable for how they fit together.
And if you are a provider, as I am
The honest reading is that being bypassed is feedback, not betrayal. Nobody deliberately excluded us. We were simply not present in the category at the moment the question came up, and presence is earned in the months before a decision, not in the meeting where it is made.
The instinct is to reach for the contract and insert a clause requiring notification of AI purchases. Resist it. A client who has to tell you is a client who wishes they did not have to. The only durable position is to be the party they would have called anyway — which means turning up with the option space, the governance position and the cost picture before they go looking. If your monthly report to a client contains uptime, tickets and patch compliance, and nothing about the AI running in their tenant, you have already told them which conversations you are part of.
The contract is the last thing to go, not the first. By the time it is in play, the decisions that mattered were made somewhere else, by someone else, months ago.
.png)
.png)
.png)