We stopped buying our software

Thoughts from a Global Citizen, edition 5. We stopped buying our software.

2016: AI before ChatGPT.

In 2016, in Moscow, I launched a startup: Trafficintellect. The idea: measure visitor flows in museums and exhibitions. Cameras counted visitors, algorithms analysed their paths, and museums could organise their rooms better.

It was the trend. Every startup was "doing AI". Nobody talked about LLMs yet: it was all machine learning and big data.

Back then, AI was a specialists' game. You needed data scientists, months of development, real budgets. For a small business, it was a distant promise.

What has changed since then isn't just how powerful AI has become. It's who gets to build with it.

A piece of advice in Los Angeles.

I discovered ChatGPT as soon as it came out, in late 2022, like a lot of people.

Then I got lucky. In spring 2024, I was passing through Los Angeles and caught up with my cousin, who works in tech. He didn't talk about ChatGPT as a gadget. He told me: "Get the paid plan. Right now."

I listened. It's probably the best tech advice I've ever been given.

For a long time, I used it like everyone else: writing, summarising, translating, thinking things through. Useful, but nothing revolutionary.

The real turning point came later, when I realised AI wasn't just for writing text. It could write code.

And here I have a confession to make: I was always too lazy to learn how to code. I understood programming logic, algorithms, software architecture. But writing the code myself? I simply couldn't.

AI removed that barrier.

The project that failed.

In spring 2026, I started my first real tech project built with AI: an internal "hub" for Axiom.

The ambition was huge. This hub was supposed to become the company's central tool: read our emails and messages, centralise everything, analyse everything.

I went in with very little method. I was discovering the tools, I wasn't using Claude yet, and I wanted to do everything at once.

The result: within a few weeks, the project was abandoned.

But along the way, I ran a test: connecting WhatsApp to a program I had written with AI. And it worked.

That's when I realised we could do far more than I had imagined.

> My first real project built with AI was a complete failure. It taught me everything.

The next lesson came quickly. For our partnership with Cours Legendre, a French distance-learning school, I built a small CRM, much more modest: one tool, one specific need, one team. It took me one day. Just one.

A tool that tries to do everything does nothing. A tool built for one specific task in your business works.

Leaving our software behind.

For years, Axiom ran on TutorCruncher, a management platform for tutoring companies. It cost us almost 15,000 euros a year.

The price wasn't the real problem. The real problem was that it was generic. Designed for everyone, and therefore for no one in particular.

The result: we were the ones adapting to the tool. Our teams worked around its limits, and our processes got twisted to fit its boxes. Nobody is comfortable working that way.

> For years, we adapted the way we work to our software. With AI, it's finally the other way around.

In early August 2026, I started building MyAxiom, our own software. This time with method, and with Claude.

Three months later: more than 1,600 recorded changes, and MyAxiom runs the company. Family requests, student follow-up, lesson reports, and since September, our tutors' payroll.

On 1 September, we officially launched MyAxiom across the whole team. And I want to give a shout-out to our education advisors: in the middle of back-to-school season, they got to grips with the tool in a matter of days and adapted impressively fast. Software is only worth something if people actually use it.

This week, a tutor wrote to Mathieu, my business partner, after filling in her details in the tool: "I find the platform really professional, intuitive and very pleasant to use. Everything is clear and easy to understand."

No software vendor knows our families and our tutors better than we do. The know-how that matters is no longer code. It's the business itself.

210 pages, then 1,940.

Same logic for our website.

In May 2026, Axiom's website ran on Tilda, a no-code website builder. Around 210 pages, all made by hand. We claimed four languages, but in reality, only French truly existed.

Today, the site is built by us, with AI. It has 1,940 pages.

The biggest difference is the public profiles of our tutors. Around 600 tutors now have their own page, in French and in English. On Tilda, that was unthinkable: 1,200 pages to build and keep up to date by hand.

These pages now come straight from MyAxiom. When a tutor is approved in the tool, their profile exists on the website.

Careful, though: "pages" doesn't mean "pages written by AI". Every profile belongs to a real tutor, with their background and their subjects. What we built is the system that publishes them.

And it's not just Axiom: the new IB Tutor website came out of the same workshop, in the same months.

What's next: augmenting the whole team.

So far, I've mostly talked about what I built. But that's not the real story.

> If AI only makes one person in the company more effective, we've missed something.

The next step is to augment every member of the team. Not to replace anyone: to augment them.

It's already happening: Mathieu, my business partner, built a "Copilot kit" for our education advisors: an AI assistant installed on their computer that helps them keep track of their messages and their families. Valentin, my brother, produces all our content with AI, from visuals to campaigns.

My ambition is simple: that everyone, in their own role, becomes ten times more effective. Not to work ten times more. So the company can grow without anyone burning out.

And now for the uncomfortable part.

Many small and mid-sized businesses still see AI as a conference topic. They pay for generic software, adapt their teams to tools designed for someone else, and wait and see.

Waiting to see is already falling behind. Those who start now won't just have better tools. They'll have teams who know how to use them.

One real question remains: data sensitivity. Handing your customer files, your conversations and your numbers to an AI is not a trivial decision. Every company has to make that choice, with eyes open.

But that choice no longer means going through the United States or China. Mistral AI, a French company, hosts its customers' data in the European Union by default. If you don't want to hand your data to an American or Chinese provider, there is a European option.

Data sensitivity is a reason to choose your tool carefully. It's not a reason to do nothing.

I'm not saying it's easy. My first project failed. But that's exactly why you need to start early: you only learn by building.