About fifteen years ago I sat at the back of a room while a man in a very good suit told a dozen small business owners they did not need to own servers any more. Some of them got it that morning. Some argued about it for another three years. The ones who moved early are, for the most part, still trading.
I think about that room a lot at the moment.
The same play is running again, and the shape of it is almost funny. A thirty-person firm in 2026 can sit one curious member of staff in front of a tool like Claude Code, on the 4.8 model, and get work out of it that used to need a whole team, a real budget, and a wait measured in seasons. Not a toy. The actual job. Someone fairly junior can now read a code base they have never seen, draft the change, write the tests for it, then explain the whole thing back to you in words you can take into a board meeting. The monthly cost is less than the firm spends on coffee.
That is the cloud promise, word for word. You no longer have to own the big expensive thing to get the use of the big expensive thing.
Here is where the small firms will get it wrong, the same way a lot of them got cloud wrong. They will treat it as a way to cut a cost, when it is really a way to build a skill. The owner reads that a model can do the work of three people, does the sum, and starts thinking about who to let go. That is the 2010 mistake again. The firms that won with cloud kept their people and pointed them at bigger work, because the dull part had suddenly got cheap. The tool itself is not the advantage, whatever the sales deck says. The advantage turns up later, in what your people do with the hours it hands back to them.
So my whole strategy note for a small business in 2026 fits on a beermat. Pick one real problem that is costing you money. Put your best, most curious person on it with the tool. Give them a fortnight and a bit of cover from the day job. Then look hard at what comes back. You will learn more from that than from any report, this one included.
The big firms have the other problem and the same root.
They are looking hard at AIOps now, which is a grand way of saying you let a model sit and watch your logs, your alerts, your three in the morning incidents, and the wall of noise nobody has had time to read since 2019. The promise is a good one. The model sees the pattern before the pager goes off. But the board meets and asks the question they asked about plain compute back in 2014. Does the data go out, or does it stay in. Cloud, or the floor below us.
For a while that was an easy call. The cloud was where the clever tools lived, so out the data went. It is harder now. A bank or a hospital trust has rules about where a patient record or a trade is allowed to sit, and a model that reads your incident data is reading some of the most telling data you hold. The new part, and it really is new, is that running a strong model on your own kit has gone from a science project to a real choice in about eighteen months. So the honest answer to cloud or on premise for AIOps is the boring one. It depends on what the data is and who is allowed to see it. And on how fast you need an answer when something is actually on fire. Anyone selling you a rule that fits every firm is selling you something.
The part nobody is paying for, again, is the people.
We did exactly this with cloud. Firms bought the platform, moved the workloads, ticked the box, then quietly noticed that almost nobody on staff knew how to run the new thing well. So they hired in a panic, paid over the odds for the few who did, and let everyone else feel a bit slow. Ten years on, the same firms are lining up to repeat it with AI. They will buy the seats and forget to teach a single person how to think with the tool in their hands.
Reskilling for this is not a half day course on how to write a prompt. It is closer to what the better firms did when they moved to cloud and actually meant it. You give people real problems and real time on them, plus the room to be slow and clumsy for a while. You accept that your sharpest database person might be average at this for six months and then, one Tuesday, very good indeed. You treat the learning as the work, rather than as something you get to once the work is done.
That is the bit that worries me, watching it all kick off again. The tools will be fine. They are already better than most of us can use. The cloud years told us where this ends up. The firms that won were the ones who spent on their people while the rest were still in a meeting about the kit. We have the same choice in front of us now, and about the same amount of time to make a mess of it.