One of my board advisers made an argument last week that I haven't been able to shake. The longer organisations wait, he said, the easier AI will become to deploy, and the less they will need a third party to help them do it. "We've seen this repeatedly with Microsoft products."
He's right. I've watched that cycle for twenty years. And I share the instinct: the same week, I told my own leadership team we have about twenty AI builds sitting half or 80% done, and that with so much "just around the corner that might change everything", I didn't want to put too much more effort into finishing them.
But he's only half right. Waiting is a sound strategy for building AI. It is a poor strategy for running it: knowing what it costs, who can switch it off, and what it replaced. Most mid-market organisations are treating those as one decision.
The build really is collapsing. My CTO summed it up in five words: "anyone can do anything now." A technology leader at a ten-hospital veterinary emergency network told us he could "ask an AI itself and figure out the steps and create an agent." Last week, for fun, I turned a client's how-to guide into a training video. It took about two minutes of my time.
The evidence on waiting supports my adviser too. IDC found that small and mid-sized businesses that held back as "fast followers" let larger firms absorb the cost of failed experiments, and are now switching on AI that arrives inside the tools they already use.
But listen to what that veterinary leader said next: "The point is, how do we actually operationalise it? How do we actually intertwine it with our licensing models so that we don't pay a huge amount of money?" Building wasn't his problem. Running was.
They assume the hard part of AI gets easier at the same rate as the easy part. It doesn't. What gets cheaper every quarter is making an agent. What doesn't get cheaper is knowing what it touches, what it costs at volume, who can switch it off, and whether the process it replaced was worth automating in the first place.
That knowledge only builds up by running something. When one AI tool moved from free to pay-as-you-go recently, the first reaction I heard of was to switch everybody off it. Nobody had a number for what it was worth, so the decision was easy. A prospect told me that version control and source files are, in his words, "all gone" for the apps business users now build for themselves. Someone will have to own those apps on the day they become critical.
The wider data says the same. McKinsey research suggests around 62% of organisations are experimenting with agents, but only about 23% have scaled them in at least one function. Experimenting is cheap. Operating is where the gap is.
My adviser's argument is a direct threat to any provider whose AI offer is "we'll build it for you". If that is your provider's pitch, waiting makes sense: the tools will soon do what they do.
What doesn't become a commodity is the operating discipline: the runbook, the cost model, the before-and-after baseline, the decision about what to retire. A managed service provider installs and maintains. What I call a managed intelligence provider has to run the thing and answer for what it cost and what it changed. That is a standing job, not a project, and it gets harder, not easier, as the number of agents grows.