22.10.2026 — Summary. AI entered companies through their employees, without a framework. nuclai, founded in autumn 2026 by Sofia and Pedro Mestre, brings measurement: knowing where to start. After a review of the organisation, we show what can be automated or handed to AI agents under human supervision, then deliver a structure the client owns.
Updated on 22.10.2026
In many companies, AI did not wait for a management decision. It came in with the employees themselves, one tool at a time, one desk at a time. Nobody did anything wrong. But nobody measured anything either: the company does not know who does what, how, or with which tools, and the productivity gains from AI are recorded nowhere. We are founding nuclai to close that gap.
How did AI get into companies?
Through the side door. Each person adopted the tool that helped with their own task, at their own desk. Strategic projects came later, if they came at all. The result: scattered tools, generic licences bought without an overall plan, and confidentiality gaps that no one has mapped.
That is not a criticism. It shows that teams want to move forward. But usage that spreads without a framework ends up costing: in licences, in lost time, in risks to data.
One example, from experience, on a question companies ask all the time: does a free ChatGPT account use the data you put into it to train its models? A company director who had “trained” her staff in AI told me it did not, as long as the project memory was switched off. That answer is wrong. By default, conversations in a free or Plus personal account may be used to train OpenAI’s models, and turning off memory changes nothing. Only the “Improve the model for everyone” setting (Settings › Data controls) prevents it, as do temporary chats (source: OpenAI). The inner workings of large language models (LLMs) have become so complex that understanding them now takes real experts, dedicated to that field alone.
What do companies really lack?
Not motivation, and not technology. Teams are ready and the tools are available. What is missing is measurement: knowing which workflow consumes time, what data flows through it, where AI will have a real effect and where it will have none. Also missing: agreement on which tools the company authorises, their rollout with secure accounts, and the sharing of working methods among colleagues in the same department.
We founded nuclai on that observation, shared by two careers that seem to have nothing in common: finance and communications. In both, we saw the same lack of clarity, and employees moving at speeds that keep drifting further apart. On one side, people who still sort the department’s tickets by hand; on the other, people who have automated that low-value task with Claude and clear it in a few seconds. Good practice has to spread, and management has priorities other than learning this new trade. That is where we come in, as specialists. Our line fits in two words, “In focus”, and one promise: “less noise. more signal.”
Who are we?
Sofia Mestre started out as a technology journalist. She then spent nine years as a project manager at a Swiss language service provider, where she worked with neural machine translation engines in production, developed both in-house and by outside providers. She holds a diploma as an AI specialist for marketing and communications and SAWI’s “AI Practitioner” certificate. She leads the communications practice: nuclai / comms.
Pedro Mestre is a trained accountant and has worked in finance since 2013. His career runs from Swiss SMEs to analytics leadership at a technology company. A specialist in commission management, he has used LLMs every day for more than two years to run the accounts payable, fixed assets, purchasing and commissions departments. He leads the finance practice: nuclai / finance.
Two areas of expertise, one standard. We work from the canton of Bern, in French and English.
How do we work?
We read before we build. Every engagement follows the same four steps, in finance and in communications, and at the end of each one the client decides whether to go on.
| Step | What we do | What the client receives |
|---|---|---|
| 1. Diagnostic | We read the process as it actually runs, on site, with the people, their screens and their files. | A map of the process, a ranked list of possible AI uses with gain, cost and risk, and what is better left alone. |
| 2. Pilot | We build a single workflow inside the existing tools, running in parallel with the old process, with a human check on every output. | A workflow that runs on its own data, compared with the starting point where the data allows. |
| 3. Rollout | We add the next workflows one by one, in the order set by the diagnostic. | A system that grows at the pace the client chooses, and that the client can stop at any time. |
| 4. Handover | We train the team on its own cases. | Instructions, calculation models, glossaries, data flows: a documented system and a self-sufficient team. |
The first conversation, held remotely, is free: its purpose is to define the scope. The diagnostic is paid, because it is the work itself, not a sales pitch. It exists for both our practices; in communications, it takes the form of a diagnostic & audit.
What if the right answer is to automate nothing?
Then we put it in writing. A process that is already efficient, too unstable, or held back by an organisational problem gains nothing from automation. Automating the wrong task only speeds up a wrong result. You can count on this: our goal is not to sell features.
A reasoned “no”, backed by evidence, is a valid conclusion. The client keeps the map and the reasoning, with no sale forced on them. It is up to the client to decide how far to take automation. If a plan is rolled out, it belongs to the client, who can entrust it to nuclai, to their own team or to another provider.
Why must the structure belong to the client?
Because a system you cannot run on your own is not a tool. It is a dependency. We hand over everything needed to use it without us: a prompt library, skills, a user guide and the logic behind each step. The system works with the language model the client chooses, or with several, and can switch whenever the client asks. Nothing stays locked in, neither with us nor with an LLM vendor. One promise matters to us: the client keeps sovereignty over their processes and can use them elsewhere. Hence detailed documentation, usable with whichever providers the client chooses.
That sovereignty is also the starting point for AI governance: knowing which tools are in use, with which data, under which rules. Switzerland has no AI-specific law. AI falls under existing law, notably the Federal Act on Data Protection (FADP), the Code of Obligations (CO) and the Federal Act on Product Liability (PrHG). On 12 February 2025, the Federal Council decided to ratify the Council of Europe Convention on AI through targeted amendments to existing law, complemented by non-binding measures; a draft is due to be put out for consultation by the end of 2026 (source: Federal Office of Justice, in French; press release in English). For companies that want to set an internal framework, as a basis for what comes next, we offer support with AI governance. This article is not legal advice.
What do we hope to bring to companies?
Clarity, and a calm adoption of AI that serves efficiency gains. We impose no moral vision: whether the client aims to cut costs, scale up or uphold an internal ethic, we apply their strategy, in complete confidentiality.
In Switzerland, as everyone knows, steady work opens doors. We live by that principle and share it with the companies we support: serious effort, placed where it creates real value.
To see how an engagement runs, from the first conversation to the handover, read our approach.
By Sofia Mestre