Last week, my team nearly walked past the biggest AI opportunity in a customer's business. It wasn't in their Microsoft 365 tenant. It was in a shed, and in the heads of two people.
"It's two women, and it's a black art."
That's how the IT leader at an agricultural processor described the process that decides which stock gets blended into which customer order. The two people know the grades. They plan the blend days in advance. And, in his words, "there's absolute treasure in that for us": thousands of tonnes of lower-grade stock that could be blended and sold instead of downgraded.
Then he said the part I keep coming back to. "We want to call it AI, but it's a freaking algorithm."
Sit with that for a moment. The plan that decides what gets sold and what gets downgraded lives in two people's heads. Nobody has written it down because nobody has ever had to. And if either of them retires, the black art walks out the door with them.
He's right on both counts. The biggest AI opportunity in his business isn't in Microsoft 365. It's in the shed: in warehouse data, in quality attributes on every bin, and in the judgement of two people who have never had to write it down. And the tool that unlocks it is probably an optimisation model, not a chatbot.
Meanwhile, by his estimate, "90% of anyone using an AI in the company is trying to write a better e-mail."
When the conversation turned to the immediate problem, securing Copilot and an AI model that had already found its way into their environment, our side of the call relaxed. Securing the tenant is what we know. As we said on the call, that's "an easier one". The blending problem got a polite "I'll have to look into that for you." In the transcript you can feel the gear change: sure-footed on the part that protects the tenant, unsure on the part that could change the business.
I'm not criticising my team. We did exactly what our industry is built to do. Managed service providers live inside the tenant: identities, permissions, devices, licences. So our AI offers start and stop there. One channel guide this year calls the AI readiness assessment "the fastest way for an MSP to take ownership of the AI conversation". It's fast because it's familiar. In our own sales review, my cloud strategist put it more bluntly: "everybody in the market is going down a security angle with AI."
The result is a quiet mismatch. The provider sells the guardrail. The value sits on the other side of a boundary the provider never crosses.
A resources company told us the same week that a lot of its AI will arrive inside vendors' platforms, which "run their LLMs in the background". Someone in his business had asked, "instead of getting a data lake, can we do agentic AI?" Both live outside the tenant: in operational systems, supplier contracts and data that Copilot will never see.
Industrial vendors are already talking about retiring experts and lost know-how on the factory floor. What I don't hear is the managed service provider, the one partner already trusted with the keys, offering to go and find it.
We may not be the people who build the blending model. But a Managed Intelligence Provider should be the one who finds it, connects it to the data, and keeps it running safely once it exists.
So here's my question for you. What's in your shed? Who are the two people whose judgement quietly moves your margin, and has anyone ever asked them how they do it?
Governance protects the tenant. The treasure is in the shed.