8 November 2026
Why my AI solutions always have a person who checks
There is a temptation, when you build an artificial intelligence tool, and it is to let it go. The demo looks better if the answer leaves on its own, in a second, with nobody touching it. I refuse that temptation. In the solutions I design there is always a person who checks the result before it becomes a decision, a document, a message to a customer, a number someone will use.
A line of responsibility
This is not caution as a slogan. It is a line of responsibility. If the tool is wrong and nobody looks, the mistake still has my name on it, because I put it in the middle of the work. If someone looks, the mistake can stop. The time you “lose” checking is the time you do not lose redoing, apologising, rebuilding a file that started crooked.
I learned to distrust things that leave without a human step long before I worked on AI. In a shop, a receipt that does not add up does not fix itself. In a company, a number in a report that nobody reread becomes, the next month, a decision. When at To Be I was putting procedures and reporting tools in place for energy projects, the point was not to produce more paper. It was to make a step repeatable and to make it possible for someone to verify it. An SOP nobody reads is a file. An SOP with a check is a trade.
A confident sentence is not a checked sentence
With AI the risk is higher, not lower, precisely because the form is convincing. A well-written sentence looks checked. It is not. Imagine — and this is a hypothetical, not an incident I am telling about a client — a firm that asks an assistant “which version of the estimate do we send?”. The tool picks one, in a confident tone, because in the pile of PDFs that version has the latest file date, not the latest decision. An email goes out. On site, people work from an old number. The thirty seconds saved cost a variation. No model “knows” which version is right if the rule was not written and if a person does not confirm the doubtful cases.
Three columns, not two
That is why Ask the Boss, the example I use on the site, has three columns and not two. The team asks. The assistant proposes. A person validates. The simulation is not a client and it has no measured result: it is the flow I want to be seen. Without the third column, the assistant is a faster switchboard, not a help. With the third column, the owner can leave the questions that have already been answered and stay on the ones that actually change something.
Checking does not mean rereading everything every time as if the tool did not exist. That would be double work, and then it would be better not to build it. It means deciding, while mapping, what can pass with a quick check and what must not pass at all. An internal reminder is not a set of accounts. A text that goes to a customer, or a number that enters an official document, does not leave on its own. The list I write about in the other note is for this too: to separate.
Trust
There is a question of trust, and I do not treat it as a conference theme. People use a tool if, when it is wrong, there is a way to stop it and if they know who answers. If the answer is “the machine said so”, they will not use it for serious things, or they will use it and then pass the blame around. I prefer a short circuit: the tool proposes, a person with a name confirms, and if needed the rule is corrected. Trust accumulates that way, not with a demo.
I have led companies where the numbers mattered: as CFO at HKR System and at Phytos Nutraceutica, two startups I also founded, and as chief executive and general manager at Universal. I am not saying we used the agents I talk about today. I am saying the reflex is the same. A number nobody has looked at is not a number. It is a hypothesis in a cleaner typeface.
The service Atlas is preparing — I write about it separately, with the link to the company — keeps the same shape: documents come in, an agent extracts, the software prepares, a professional validates. If you remove the last step, what remains can be fast and unreliable. For an energy diagnosis, a carbon footprint or a sustainability report, unreliable is worse than slow.
I do not promise that the check makes everything safe. I promise that without it I do not sign the work. The free thirty-minute review, if we do it, includes this question too: who, in your company, is the person who can say “yes, this can go out”? If there is no answer, we do not start from the tool. We start there.