How Takeda France trained three quarters of its head office on AI

Three years ago, a small team at Takeda France started an internal programme to teach colleagues how to use generative AI. It was called ExplorIA Academy. It is now in its third season, it has been run throughout by two people who have other jobs, and by the December 2025 checkpoint, 75 percent of head office staff had attended at least one session.

Attendance was voluntary throughout. Nobody was required to take a Copilot licence, and nobody was required to show up.

MedShake Studio has worked on ExplorIA since its first season. The programme is Takeda's, designed and run by Aurélie Suchaud, Learning and Development Partner, and Johanna de Silva, NGCE Business Partner. Our part has been the one a two-person team cannot hold alone: co-designing each season's architecture, building the session content, watching the field so the roadmap stays current, and more recently building the tooling that lets the programme scale past its two founders. Suchaud put the reason plainly in the interview: generative AI moves faster than two people with other jobs can follow.

The figures below are Takeda's own, presented by Suchaud and de Silva in an interview recorded for Tech Anatomy in June 2026.

Three seasons, three different problems

The seasons were not iterations of the same course. Each one addressed a problem the previous one had left open.

Season one was about fear. When Johanna de Silva joined, the obstacle was not access to tools, it was a set of beliefs: that AI would replace people, and that the subject was too technical for anyone outside data and technology. The season was built to demystify, and it leaned on gamification to get people through the door. Its content was organised by job function.

It also ran into a constraint specific to the industry. Takeda restricts the use of external software, so the tools shown in season one were often tools employees could not then use at work. The programme could change how people thought about AI, but not what they did on Monday.

Season two started from the opposite end. In May 2025 the company obtained Microsoft Copilot licences, and Aurélie Suchaud, who had joined at the start of that season, made the licences the backbone of the year. Sessions were built tool by tool, starting with Teams and Outlook because everyone uses them, and about forty concrete use cases were demonstrated across the season.

Season three, which is running now, is about handing the programme over.

The choice that shaped season two

Running sessions tool by tool rather than function by function is not the conventional approach to corporate AI training. The usual pattern assumes people will pick up the basics alone and puts the training effort into what is specific to each job.

Suchaud and de Silva were explicit that the conventional approach is the better one in principle. In their words, the best of all worlds would be sessions built by job function. They chose the other path for two reasons. The tool was new to everyone, which meant an unusually level starting point and a chance to give the whole company a common base. And there were two of them, part time, which made a session catalogue segmented by department impossible to staff.

The choice had a visible effect on demand. Sessions were capped at 30 participants, then raised to 40, and filled within 24 to 48 hours. Sessions had to be duplicated.

It had a less visible cost too, and it is the natural question to ask of any programme built for the widest possible audience: content pitched to be accessible to everyone is content that some senior functions will find below their level. The team's answer for season three is agents, which are specific enough to be worth a specialist's time.

The measurement, and what it does not measure

The 75 percent figure is an attendance rate: the share of head office employees who came to at least one session, on a voluntary basis, measured at a December checkpoint. It is a strong number for a voluntary internal programme, and it is a measure of interest rather than of proficiency or of sustained use.

The team is straightforward about this. Their other evidence is of the same kind: sessions filling in under two days, sessions duplicated to meet demand, managers in the field asking for the programme to reach their teams in 2026.

One measurement in the sessions did point at something operational. A poll opening the Teams session asked how many hours a week people spent in meetings. The answer came back between 15 and 20. The use cases that followed, automatic meeting minutes, questioning a meeting after the fact, turning notes into an audio summary, landed against a number everyone in the room recognised. Marketing teams reported translating slide decks by prompt as the single change they valued most.

What nobody has yet is a measure of hours actually saved, or of how many of those practices survived six months. That measurement is harder to produce than an attendance rate, and it is the one that would settle whether the programme changed behaviour or awareness.

Season three, and the limits of two people

The constraint that shaped every season was capacity. Two people, neither full time on the programme, cannot be present in every department, and cannot rebuild a roadmap every time the technology moves. Suchaud described revising the 2026 roadmap in April, then revising it again the following day because something new had to go in.

Season three answers this structurally rather than by working harder. An ambassador community was created a few months ago, one volunteer per team, with the intention of canvassing department by department to fill it. The canvassing turned out not to be necessary and the community now numbers around thirty people.

Their brief is not to relay announcements. It is to bring their team's actual needs back to the centre, so that agents can be built against real workflows rather than generic ones. The rituals are deliberately light: a monthly meeting, an informal half hour on Fridays, a Yammer channel, and Coffee Chats open to the company where each ambassador presents one of their own use cases rather than the programme presenting it for them.

The stated intention is that ExplorIA stops being Aurélie and Johanna's project.

Season three is built on four pillars that reinforce each other rather than four separate initiatives. Talks, ninety minutes in the cafeteria where internal pioneers and outside experts present, to inspire. Coffee Chats, a short monthly briefing on what has changed in AI, to inform. Training sessions on agents, Copilot assistants and advanced prompting, now segmented by job function, to equip. And the IA Champions community itself, with its Yammer animation and peer support, to sustain.

The tooling followed the same logic. Rather than build agents for every team, which would recreate the bottleneck the community was meant to remove, we built L'Architecte, a Copilot agent whose only job is to write the instructions for other Copilot agents. Someone describes in one sentence what they want to automate, and it returns complete instructions ready to paste, with the associated parameters. It takes fifteen minutes to deploy and requires no prompt engineering knowledge. The point is not the agent, it is that a team of two can now say yes to a department without writing the agent themselves.

What they would tell someone starting

Asked what to avoid, de Silva answered without hesitation: do not rush. The ambassador community was on the whiteboard from season one, and building it then would have meant recruiting into a programme that had not yet generated any enthusiasm to relay.

Asked what to do, both gave the same answer: stay practical. Not a curriculum about artificial intelligence, but a demonstration of something the person in the room does on a Tuesday, done faster.

Five things worth stealing

Most of what made ExplorIA work is transferable, and none of it is expensive.

Anchor the first session in a number the room already recognises. The Teams session opened by asking how many hours a week people spent in meetings. The answer, 15 to 20, made every subsequent use case land against something everyone in the room was already living. A training that starts from a capability is a training about software. A training that starts from a number people recognise is a training about their week.

Pick the sequencing your team can actually staff. By job function is the better design and nobody disputes it. By tool is the design two part-time people can deliver to a whole head office in one season. The right sequencing is the one that ships, and the tool-first path also produced something unplanned: a common vocabulary across departments that function-specific tracks would not have created.

Make attendance voluntary and let demand do the arguing. Sessions filling in 24 to 48 hours and needing duplication is a stronger argument in front of a leadership committee than a completion rate obtained by mandate. It also tells you where real appetite sits, which a mandatory programme cannot.

Build the community after the enthusiasm, not before. De Silva's own answer to what she would avoid: do not rush. The ambassador community was on the whiteboard from season one. Built then, it would have been recruitment into a programme with nothing to relay. Built two seasons later, it filled itself without canvassing.

Give ambassadors a job that produces something. Relaying announcements is not a role, it is a mailing list. Bringing back their team's actual needs so that agents get built against real workflows is a role, and it is the thing that survives the initial goodwill.

If you are building something similar, this is close to what our AI adoption programmes do: the audit that establishes where a company actually stands, the academy format that ExplorIA is the longest-running example of, and the agents built for teams once the community starts surfacing what they need.

What to watch

Two things will say more about ExplorIA than any attendance figure.

The first is the field rollout in 2026. Head office and field teams do not share a working day, a device or a set of tasks, and a programme designed in and for head office does not transfer automatically.

The second is whether the ambassador community produces agents that people keep using. Thirty volunteers signing up is a measure of goodwill at the start. Thirty ambassadors still surfacing needs and presenting use cases a year from now would be a measure of something else, and it is the part of season three that has no precedent in the two seasons before it.

Listen to the episode

Le retour d'expérience de Takeda - Aurélie Suchaud & Johanna de Silva

Listen · 24 min