Everyone has a view on AI now, and I've been wary of adding to the pile. Most of what I read about it is written from the outside, by people who haven't had to make it work on a Monday morning, with real customers on the other end and a budget that has to be found from somewhere. So rather than post about it, I've mostly been doing it, and writing down only the parts that held up.
I should be clear about what I know and what I don't. I'm not a researcher and I don't build models. What I know is what happens after the demo. I've spent my career taking new technology into businesses, the web, then mobile, then data, now this, and finding out each time which part reaches the accounts and which part was a good demo. The pattern is always the same. The capability arrives on schedule. The value turns up later, somewhere slightly different from where everyone was looking, and only in the businesses that did the unglamorous work in between.
Three things have held up well enough that I'm prepared to be held to them.
What I know is what happens after the demo.
The first is that when the work gets cheap, a business is worth its judgement. Every competitor can rent the same models, and the models get better and cheaper every few months. What nobody can rent is the people who can tell a good answer from a plausible one, the standards you've written down, and the record of decisions you've taken. That is the layer worth owning, and most AI programmes don't touch it.
The second is about people, and it's the one that worried me first. A lot of what AI does well is the work juniors used to learn on. Take that away without rebuilding the learning, and in ten years you have nobody who can supervise the machine. I wrote this up as a prediction with a date on it, so that in a few years someone can tell me whether I was right.
The third is that an agent should earn the right to do more. It starts by watching, then recommending, then preparing, and only acts on its own once it has met criteria written down before it started, with a named person who answers for it. I've found this is the thing boards actually want to hear, because it answers the question underneath all the others: who is accountable when the machine is wrong.
There's a distinction running through all of this that I came to late. Putting AI into the work a business already does and deciding what a business should sell next are different jobs. The first needs tight controls. The second needs room to try things, and a way of charging for the outcome rather than the hour. I've spent more of my career on the second than the first, and I think most of the value is there, once the first is under control.
All of it is on a site of its own, ai.valuecreator.io, because it's a working set of things rather than a set of opinions. There are pieces organised by the questions leaders actually ask me, a five-minute check that gives you one page for your next leadership meeting, and the playbook I use, which you can copy. There's also a Claude plugin that runs the exercises with you. I made it follow its own rules, so it has a test set and a register, and it says in public which of its own tests it still fails. I change my mind in the open too; the About page keeps a log of what changed and why.
None of it is a product and I'm not selling anything. It's the way I work, written down so other people can use it and tell me where it's wrong. If you run one of the tools on something real and it gets it wrong, I'd like to know. I'd rather be corrected by someone who has done the work than agreed with by someone who hasn't.