Stopping Anatole

Anatole was a conversational assistant for tourism offices. It spoke to territorial databases, drafted content, answered visitors. Technically, we were around 99% of answers rated qualitative by our testers. First clients were signing. The team believed in it. So did I.

And yet, early 2025, I shut Anatole.ai down. Not out of exhaustion. Not out of a lack of traction. Out of lucidity.

Shifting ground

The first reason is something nobody wants to look at directly. The technology isn't ready, and the gap between base models and vertical products shifts every three months.

In eighteen months, we rebuilt Anatole three times. Three full rebuilds to keep up with ChatGPT, Claude, RAG tooling. Each rebuild was weeks of work that a single model update made redundant overnight.

On October 31, 2024, OpenAI rolled web search into ChatGPT.^[1] From one day to the next, features we had spent months building became useless. Our cross-verification layer, our RAG on tourism data, our citation module: all replaced by a single update from the vendor of the model we were running on.

Building a vertical product on a substrate that shifts every quarter is coding on sand. You can do it once. You can do it twice. By the third time, you have to ask whether the sand is ready to carry a house.

The 1% that lies

The second reason is statistical. It still keeps me up.

An AI that answers correctly 99% of the time is an AI that lies once in a hundred. In casual contexts, it's nothing. In mountain tourism, it's the difference between a reassured visitor and a body to search for.

A ski touring recommendation that ignores the day's avalanche bulletin. A pass listed as open when it's closed for roadwork. A last cable car time off by an hour. These errors weren't theoretical. We ran into them in tests, quietly, in the margins.

Here's the thing. The 99% of correct answers make the 1% invisible. When almost everything is right, users drop their guard. They stop checking. That's exactly when the rare error does the most damage.

High reliability in casual contexts is a benefit. High reliability in critical contexts is a trap. Those aren't the same skill.

And two years later, the problem hasn't gone away. GPT-5.4, OpenAI's most recent model as I write this, still clocks over 10% factual hallucinations on summarization tasks, according to the Vectara benchmark.^[2] The models marketed as the smartest, the ones that reason, hallucinate more than simpler models on basic factual recall. Progress on reasoning is real. Progress on factuality in critical contexts is slower than the ambient noise suggests.

Who pays when the machine lies

The third reason is legal, and it's the coldest.

In February 2024, a Canadian tribunal ordered Air Canada to refund a passenger whose chatbot had falsely promised a retroactive bereavement fare.^[3] The airline's defense was that the chatbot was a separate legal entity. The tribunal dismissed the argument in one sentence: it should be obvious to Air Canada that it is responsible for all information on its website, whether that information comes from a static page or a conversational agent.

In the European Union, the framework is tightening. The Product Liability Directive, revised in 2024, now classifies AI software as products.^[4] To obtain compensation, a victim no longer has to reverse-engineer the algorithm. They prove damage and defect. The burden of proof flips. The AI Act, rolling out in phases since August 2024, layers its own obligations on high-risk systems.^[5]

I kept asking my team the same question in private. When Anatole gives a bad trail recommendation to a hiker who gets injured, who's liable? OpenAI's model? The team that built the RAG? The tourism office that signed the contract? The founder, meaning me?

The honest answer was that nobody knew. And when in doubt, courts look at the last visible commercial link in the chain. That meant the tourism office. That meant us.

Stopping is a skill

The argument to keep going was easy. The market is moving, competitors are moving, whoever leaves first takes the share. We repeated this mantra in pitch rooms, in keynotes, in articles. Hold on, iterate, survive until the tech catches up with the vision.

Except the real risk, in an environment changing this fast, isn't stopping too early. It's continuing too long out of inertia, pride, or fear of having invested for nothing.

I shut Anatole down for three reasons that fit in three sentences. The technology wasn't ready, and we were chasing updates faster than we could code. The critical use cases in mountain tourism don't tolerate 1% error, and we had no credible path to zero. The legal framework had become clear enough that I knew the next victim of a hallucination would be our problem, not the model vendor's.

In the rooms where I train tourism professionals, I watch executives hesitate every week before launching their own chatbot. They're right to hesitate. I don't push them. I tell them about Anatole.

The values this taught me aren't slogans. Reliability before speed. Safety before performance. Humanity before automation. They came from a product that was working, and that I stopped.