A friend who works at a large Swiss company told me the scene: the AI licences are there, paid for, deployed across hundreds of workstations. And nobody really knows what to do with them. A few tries in meetings, two or three questions asked "to see", then everyone went back to their habits. We observe this pattern everywhere — in large groups as in SMEs. The problem is almost never the tool. It's what's missing around it: the integration.
A licence is not an integration
The usual order is reversed. Companies buy first — Copilot, ChatGPT Team, some assistant — and hope usage will follow. It doesn't, for a simple reason: a generic tool knows neither your quotes, nor your clients, nor the way you name things. As long as it isn't connected to your real workflows, it remains one more chatbot in one more tab.
Generic training doesn't close that gap. A "prompt engineering" workshop on a Friday afternoon doesn't survive Monday morning, because it says nothing about the team's concrete cases: the quote due before noon, the reminder for the client who hasn't paid, the Friday reporting. You learn to use a hammer by driving your own nails — not by looking at examples of nails.
What we observe:
Start with the week, not the tool
The right starting point isn't a catalogue of tools. It's a typical week, taken honestly: what repeats?
Generating quotes. Chasing unpaid invoices. Answering routine emails. Entering invoices. Assembling the reporting. The test is simple: if the task follows roughly the same path every time, it's a candidate for automation. The data already exists in your systems, the format is stable, the trigger is known.
And then there's the rest — arbitrating a budget, deciding on a hire, choosing which client to meet. These aren't tasks, they're judgements. They stay human, and that's not negotiable: a responsibility cannot be delegated to a machine, even when the machine sounds convincing.
We've made this sorting exercise something you can try on our AI integration for companies page: a workbench where you sort a week's tasks between what goes to the machine and what stays with you.
Observe, install, transmit
An integration that lasts follows three steps, in this order.
Observe. Spend time inside the organisation before connecting anything. Watch how quotes actually get made — not how the process says they get made. This observation time almost always reveals two things: automatable tasks nobody had thought of, and "AI needs" that aren't.
Install. The repetitive only, with a human check at the end. AI prepares the quote, you approve the sending. It drafts the reply, you sign. This isn't decorative caution: it's what makes a machine error a corrected draft instead of a sent email.
Transmit. Train the people who will live with it — on their cases, not on examples. An automation nobody understands internally is debt: the day it breaks, the day the provider is gone, the day the process changes, it turns into a black box nobody dares to touch.
The data question, before everything else
Before connecting a model to anything, one question must be asked — and written down: what leaves, what stays?
Source by source: client contacts, internal documents, accounting entries, emails. Nothing should leave by default. Every source that feeds an external model is an explicit decision, not a factory setting. Where possible, require hosting in Switzerland or the EU. And above all: require the scope to be written down in black and white by the provider. A verbal promise about data only binds the person listening.
What doesn't need AI
It has to be said, because few vendors will say it: a share of the problems presented to us as "to be solved with AI" are solved with a well-built form, a naming convention, a shared checklist or a tidy folder.
AI applied to a blurry process produces blur, faster. The useful order is: clarify the process first, then — only if the repetition persists — automate it. Sometimes clarifying is enough. That's good news: it costs far less than a licence.
Where to start, concretely
It's the method we've applied in our own workshop for two years, and the one we now offer to SMEs in French-speaking Switzerland — the detail is on our AI integration for companies page.
Key takeaways
An AI licence is not an integration: as long as the tool isn't connected to your real workflows, it remains one more tab.
The right inventory starts from the week, not the tool: what repeats is a candidate, what requires judgement stays human.
Three steps, in order: observe before installing, install the repetitive only, transmit before leaving.
Data first: nothing leaves by default, and the scope is written down in black and white.
Many processes don't need AI — a form or a checklist sometimes solves the problem, for far less.