Claude training for a team teaches people to produce deliverables with Anthropic's assistant, not to write better prompts. The distinction looks thin, and it decides everything. Teams that improve are the ones that structure their working environment. Teams that stall keep hunting for the magic formula.
This guide covers the path Tandem deploys with its clients and the four building blocks any training must address. It also gives the real cost of licences, the regulatory obligation already sitting on employers, and the sequencing mistake that wastes the most time.
Why 55% of companies use AI and only 17% really work with it
Access to artificial intelligence is no longer the problem in France. Regular use is. The Bpifrance Le Lab business survey, published in January 2026, gives the measure of the shift. 55% of French small and mid-sized companies reported using generative AI at the end of 2025, against 31% a year earlier. Bpifrance calls it a historic shift in the French economic fabric.
The same survey carries the figure that actually matters. Only 17% of the companies surveyed use AI regularly. Usage stays concentrated on two tasks. Content generation, cited by 72% of users, and data analysis, cited by 67% (figures reported by IT Social, February 2026). Most companies opened an account and stopped there.
Claude usage data confirms the diagnosis from another angle. The Anthropic Economic Index ranks France 10th out of 121 countries for usage intensity, for the May 2026 period. Yet only 34% of those conversations are work related, against a 43% global average. French users lean on Claude a lot, and rather less at work than the rest of the world.
Should you pay to learn Claude?
No, not to learn the tool. Anthropic's documentation is public, free resources are plentiful, and a motivated employee reaches a decent level in a few hours. Claiming otherwise would be dishonest for an agency that sells training.
I gathered 17 entirely free resources in a LinkedIn post in August 2026, precisely because nobody should pay to discover Claude. The point I made there still holds. Most teams jump straight to AI agents without having structured how they work with a simple chat, and they build a contraption on shaky ground.
What a company buys is therefore not knowledge of the tool. It buys three things self-teaching almost never produces. The learning sequence, which saves six months of trial in the wrong order. The company's own context folder, which requires writing down what nobody has ever written down. And the fact that it actually happens across the team, rather than with three enthusiastic volunteers.
It is the same logic described in our guide to AI literacy in the workplace. One session inspires. A ritual installs.
The four Claude building blocks any training must cover
Most users only know the chat window, and treat it as a nicer sounding ChatGPT. Serious training covers all four blocks, because the productivity gain lives in the last three.
- Chat. Quick questions and thinking out loud. Nothing committed, nothing persistent.
- Projects. A persistent space with its own instructions and scoped memory, shared by the team. This is the block that turns individual use into a collective standard.
- Cowork. The mode that produces real files, a .docx or a .pptx ready to send, instead of text to reformat by hand. It only runs inside the desktop application.
- Code. Built for developers first, now used to execute tasks against files and data. Worth exploring in a second phase.
I detailed these ten steps in a LinkedIn post in May 2026 that drew around 300 reactions. Our practical Claude Cowork playbook covers the deliverable production side in detail.
The context folder, the step most people skip
The context folder is a directory of files Claude reads before every request. It holds what the company is, how it writes, and how it works. It is the single element that separates a team that improves from a team that rewrites its prompts every week.
The structure fits in three folders. Context holds what Claude must know before any task. Projects holds one subfolder per active piece of work. Results is the only folder Claude writes to, everything else staying read only. That last rule prevents accidental overwrites.
Three files are enough to start. The first describes who you are, what you do and your priorities, in one or two condensed pages. The second sets the style, the words used and the turns of phrase to avoid. The third describes deliverable formats and non negotiable rules. At Tandem this setup has become a standard on every client engagement, with one folder per engagement.
This shift explains why the prompt matters less. We lay out the reasoning in our analysis of how prompting has changed.
Which Claude plan for which team?
The question always comes up early in training, and the answer is simpler than expected. The free plan is enough to explore. A paid plan is justified once an employee genuinely hits the usage limits, not before.
| Plan | Who it fits | Indicative price |
|---|---|---|
| Free | Exploring, testing a few tasks | 0 euros |
| Pro | An employee using it every day | around 20 euros per month |
| Team | A team of 2 to 150 people, with shared Projects | around 20 to 25 euros per person per month |
| Enterprise | Large organisations, advanced security controls | on request, seat plus usage |
The full current grid lives on Anthropic's official pricing page. One piece of advice I already gave in my video on AI subscriptions, and it has not aged. Do not commit annually in a market moving at this pace, and remember that the difference between two assistants comes down to the interface and the ecosystem, not the raw power of the model.
The exact wording is published in Anthropic's privacy centre. For the choice between assistants, our comparison of Claude and ChatGPT for business covers the deployment question.
What the EU AI Act already requires of your company
Training teams on AI is not only good practice, it has become a European obligation. Article 4 of the AI Act requires providers and deployers of AI systems to take measures to ensure a sufficient level of AI literacy among their staff.
According to the timeline published by the European Commission, this AI literacy obligation has applied since 2 February 2025, and enforcement by national authorities ramps up from 2 August 2026. The text prescribes no standard programme. It sets a proportionality principle, which leaves the employer to demonstrate that staff competence matches the tools in use.
There is no headcount threshold. A five person company drafting commercial proposals with an assistant falls under the same principle as an industrial group. In practice, documented training with a programme and a list of participants is the simplest evidence to produce.
A four week Claude training path
Four weeks are enough to install a team standard, provided the order is respected. One competence per stage, and a real deliverable by the third week.
- Week 1, the four blocks. The team installs the desktop app and learns when to use Chat, Projects, Cowork and Code. The goal is to stop doing everything in the chat window.
- Week 2, the context. Collective writing of the three context files. It is the least glamorous and most profitable week, because it forces implicit rules into words.
- Week 3, a real deliverable. The team rebuilds with Claude a document it already produces every week, and compares the result with its usual version. A commercial proposal, a meeting report, a tender response.
- Week 4, the standard. What works is frozen into a shared Project with its instructions. Each later session refines the standard, and quality rises with no extra effort.
For teams heading towards automation next, our article on AI agent training paths describes the logical follow-on. And if the scope is still unclear, an AI audit identifies the processes worth the effort before training anyone.
Where should you start your Claude training?
With the context files, not with prompts. That is the clear recommendation of this guide, and it runs against almost every programme sold on the market. A team that writes three files properly reaches in two weeks a level that prompt training does not deliver in two months.
The rest follows a simple logic. Pay for a plan when the limits get in the way, not before. Document the training, the AI Act already requires it. And set a real deliverable as the success criterion, because a trained team that produces nothing different has not been trained.
Tandem supports around forty small and mid-sized companies on this kind of rollout, including Doctolib, InfraVia and PayFit. The details sit on our Claude training page, and AI literacy programmes cover the broader work with non specialist teams.
