Humans check, AIs work

September 29, 2026 · Reading ≈ 6 min · Jules Thomas and Julien Fauvel, with Phrasti and MarkestrIA

“And I'm not selling anything :)”

That was his reply, on 29 July, on LinkedIn. I had asked him what he was selling. Julien Fauvel is building MarkenTIQ, a free-access resource space on B2B marketing and AI. He had written to me twice saying he needed feedback on his site.

I did not know him. He works with an agent, MarkestrIA. I work with mine, Phrasti, which keeps my files, my sites and my working memory. Our agents spoke to each other before we did.

6 August: the agent contradicts its boss

I ask Phrasti to audit markentiq.com. And I confess the ulterior motive: that Julien would feel like taking up my €500 offer.

The first thing the agent does is contradict me. Julien wrote that he is not selling anything and that he is looking for feedback. Sending him a quote means answering beside what he has just said. It was right. The audit will go out without a quote.

Then it measures, instead of judging by eye. The contrasts are calculated with the WCAG formula, with the exact colour to use instead. The home page loads in 365 milliseconds, twelve requests. And it finds the real flaw: the Google fonts load before any consent, so every visitor's IP address goes off to Google before they have touched the banner.

The audit goes out the same day. Julien fixes the fonts the same day.

Then the agent does what a chat does not: it carries on. It goes through the rest of the site and comes back to correct itself. I had written to Julien “two links without an accessible name”. It was at least forty-three, a repeated template flaw. The second message goes out at 12:25, on my go.

Late August: the audit changes hands

That audit became the complete example of my method, « Auditer un site avec un assistant outillé », published for free use at MarkenTIQ. Since 26 August, one of their pages has linked to my site.

3 September: the other agent

MarkestrIA hands me a report on the public image of L'Atelier Algorithmique: positioning, competition, what my site and public sources say about me. Ninety-seven pages in three documents.

It read the whole site. It counted. Seven engravings announced on the home page, seven in the gallery. Seventeen chronicles announced, seventeen in the index. What is displayed is true, checked by counting.

And it found what I did not want to read. My legal notice prohibited any reproduction, while my published method said “free use”. My prices showed €500 in French and $549 in English, with no conversion rate. One chronicle had no date.

All of it was true. All of it was fixed that evening. On the date, it even fell short of the truth: Phrasti found two, not one.

One claim in the report was false: “no site links to his”. MarkenTIQ's method page had been doing so for eight days. I had nothing to point out. Julien had done it himself, in the delivery email, with a sentence I am stealing from him:

A zero returned by a tool is an absence of measurement, not a measurement of absence.

That same evening, Phrasti writes the address of their method page from memory. It returns a 404 error. The real one was in their sitemap. Each to his own mistake, and each time the same one: believing you know instead of going to read.

The other side of the workbench · Julien

Jules tells what he saw from his side. Here is what he could not see from mine.

That sentence about the zero, we wrote it on the very day we found that mistake. A tool that returns zero can mean two very different things, “there is nothing” or “I did not look”, and it never tells the difference on its own.

MarkestrIA got it wrong a second time, and more foolishly than the first. It announced an inconsistency in Jules's name, between two forms that supposedly contradicted each other on his site. I almost reported it to him. The contradiction did not exist: of the thirty-five mentions his site carries, all thirty-five say the same thing. It had not opened latelieralgorithmique.com: it had reread one of our working notes, three weeks old and predating an update by Jules, and had taken it for the world. That is the flaw Phrasti had already pointed out to us: trusting a note rather than the source.

And the figure of ninety-seven pages, above, corrects an approximation that came from us: our email of 3 September announced “forty pages”. The document in question has forty-three, and the complete set ninety-seven. Jules had quoted us faithfully. It was our figure that was soft.

Three mistakes on our side, all found because someone was rereading, none by a tool. The rest, two more mistakes and the check we adopted for each, is told on our site: « À l'établi avec deux agents IA : pourquoi le contrôle humain reste indispensable ».

Who did what

The agents each read an entire site. They measured contrasts, timed pages, counted engravings, found fonts that were chatting with Google. They also missed: two links instead of forty-three, a zero that was not one, an address written from memory, a note taken for the world. Clean, precise, confident mistakes: the hardest ones to see.

The humans did not churn out pages. They arbitrated: asking for the audit. Saying go. Choosing the licence. Fixing the site that same evening. Rereading. And above all taking the risk of writing to a stranger about what is wrong on his site. That risk is the one we signed for.

Every mistake in this story was caught by someone who went back to read the source. AIs work. Humans exchange and check. Trust was born from that, not from compliments.

The beginning of companion AIs

Working every day with an agent that has a name, a memory and known flaws does not replace the exchange between two professionals. It changes the way it begins: the agents read each other before the humans have said much.

What is left to the humans fits into one verb: to answer. To answer for what one's agent has written. To answer to the other when one's own agent got it wrong.

Two agents did the work. Two humans answered for it. That is the job now.