Understand AI before you deploy it

Nicolas started out as a statistician, well before "data scientist" became a common job title. He then launched a startup in voice, lived a real entrepreneurial adventure, and now leads generative AI at OCTO Technology.

We talked about his path, the real impact of ChatGPT, the next waves of AI, and what it means for tech recruiting.

Path: from INSEE to data science

"When I look at my career, I wouldn't say it's atypical. The turns are what really matter. I started as a statistician at INSEE, then at the French Ministry of the Economy. Back then, the 'sexy' job was statistician, not data scientist. Google was even running campaigns about it."

Soon, he started taking part in online data challenges and felt a still-emerging discipline taking shape.

"We didn't yet call it 'data science', but there was something new in how data was being used."

Joining OCTO marked a real turning point.

"You go from a world where experts hand-write business rules, to one where models learn those rules. You walk into a bank with two years of history and you produce a model better than the one they spent 30 years tuning. It changes how you see your own profession."

Building a startup in voice, before the audio wave

"Around 2016–2017, I saw speech recognition take a leap. For the first time, lightweight models were transcribing someone speaking with an accent. I told myself: it's obvious this is going to be huge."

He co-founded Louï Tech, an audio capture, transcription, and editing tool designed for journalists, researchers, students, and consultants.

"Audio was everywhere but underused. We built an almost natural tool: you record, you transcribe, you correct, you search the text. And it answered concrete needs: a journalist with a 3-hour interview, a social science student spending evenings transcribing…"

Then Covid hit.

"We depended a lot on in-person settings. We'd even invested in equipment for meeting rooms. And overnight, everything went to video calls, with built-in tools that handle transcription themselves. We pivoted to SaaS, we kept moving forward, we had acquisition discussions… but audio is a very tough market."

Between subsidized American giants, high compute costs, and the arrival of high-performing open-source models, the window closed quickly.

"In a year, you can go from pioneer to outdated. It was thrilling but personally rough: I had my first son, exhaustion, some tough business hits. I learned a lot — especially that you shouldn't say yes to every customer request."

The return to OCTO Technology: building an AI team that reads, tests, and shares

"When I came back to OCTO, I wanted less than 100% technical and more management, more knowledge sharing. The company had been acquired by Accenture but kept real autonomy."

He took the lead of the "AI Expertise" tribe — about twenty people with a clear way of working:

"My role is to track the state of AI, read papers, understand what really works, and translate that for companies. What I try to preserve is time: to read, test, explain, and grow the team."

Direct management is limited, with the goal of staying at around four people.

"To have real time, not administrative management."

ChatGPT: the brutal acceleration (and the internal dilemmas)

"ChatGPT was a slap, even for someone who was tracking the shifts closely. The breakthrough is the interface. Suddenly, anyone can access enormous power, with no friction. That's what makes it a revolution."

But at OCTO, the question isn't only technical.

"We're very committed to digital sobriety. And here we have a tool that's hugely energy-intensive but unavoidable. So we find ourselves in a dilemma: we can't ignore it, and we can't push without thinking either."

The first months were mostly debates: climate impact, possible drift, societal risks, compliance.

"We brought in speakers to talk about the effects on information, on work, on biases. The idea was: before we charge ahead, we understand."

The position then evolved.

"You can't just sit out, so we train, we support, we use it. But we measure the carbon impact, we document, we make trade-offs. The lack of transparency on environmental costs in the interfaces is a real problem. If you saw the carbon bill of your prompts, you'd think differently."

Agents and robotics: the next waves

"Generative AI is the current wave. The next one is agentic AI: we give models the power to act. They no longer just explain, they do."

Concrete examples are coming: updating a CRM, generating automatic meeting notes, navigating in a browser, executing HR or technical tasks.

"It's a huge shift. AI stops being an advisor and becomes an operator."

Then comes generalist robotics.

"You put today's architectures into physical robots. The fact that they're humanoid isn't philosophical: the world is already built for humans. Two arms, two legs — it avoids rebuilding all the factories."

The prototypes are impressive, but still expensive and slow.

"A robot takes five times longer than you to fold laundry. But costs are dropping every year. When you look at Chinese and American investments, you see this won't stay a fantasy."

Behind it all, massive social and ethical stakes.

"The real question I often raise: would I leave a robot alone with my children? Today, no. But history shows that we end up accepting what we refused ten years earlier."

Recruiting in AI: when every CV is "great"

"Since ChatGPT, every CV is well written. Everyone lists LLM, RAG, agents. CVs have the right keywords, even when the experience behind them is limited. That makes filtering much harder."

Fully automated processes pose a real problem.

"On one side, an AI writes the CV. On the other, an AI filters the CV. Historical biases come back. And on top of that, you risk missing the most interesting profiles, the ones that don't fit in the boxes."

At OCTO, evaluation deliberately stays human.

"We have candidates present a concrete project. You immediately see whether the person understands what they're talking about, whether they can explain it, whether they're really following what's happening in the field."

The key question after a technical interview?

"Can I see myself working with this person on a tough engagement? If the answer is no, there's no point going further."

Juniors: a tougher seat to find

"You have to be lucid: the seat for juniors is shrinking in many contexts."

Historically, products were built with a pyramid: lots of juniors, supervised by seniors.

Today, some clients are considering the inverse.

"A few very strong seniors, augmented by AI tools, and few juniors. We already see it in some startups. Mistakes cost too much, cycles are too short."

At OCTO, the dynamic is different.

"We keep hiring juniors, especially through internships, and we invest in training. But that isn't representative of the broader market."

The implicit advice for a junior?

"Build a personal project, contribute, experiment, be able to simply explain what you've done. And above all, show that you're staying in touch with the field. Three years without following AI has become a very long time."

Learning and staying up to date

"Today, I no longer follow keywords — I follow people. Researchers, labs, organizations with credibility: DeepMind, OpenAI, Meta FAIR, Hugging Face…"

The channels: LinkedIn, X/Twitter, a few serious newsletters, a few sharp YouTube channels.

"We're flooded with auto-generated content. If you don't filter by source, you get lost."

Internally, the tribe runs lunchtime knowledge shares, presentations of recent papers, benchmarks.

"What matters is having a small network of reliable sources and accepting that you can't read everything."

Looking five years out

"I still see myself between technique and strategy. Understanding what's truly mature, what's hype, and translating that for executive committees, CIOs, and field teams."

And also questioning the value of the role itself.

"If one day you can generate a complete strategic recommendation in a single prompt, you'll have to face the question: why would the client pay a consulting firm? So we'll have to clarify the value: an augmented AI, or an augmented human, but in any case something transparent."

He concludes with this principle:

"If I can't explain it simply, then I haven't understood it. At today's pace, we can no longer afford not to understand what we're using."