Technological Sovereignty: How to Deploy AI Agents Without Depending on One Tool

In September 2025, Microsoft added Anthropic's Claude models to Microsoft 365 Copilot, alongside OpenAI's. In June 2026, Copilot Cowork reached general availability running on Anthropic models. Even the tool most entrenched in large enterprises swaps engines mid-flight.

For a company, technological sovereignty is therefore less about picking the right tool than about what it keeps when the tool changes. While your teams used AI to chat, switching tools cost almost nothing. Once they build agents, skills and connections to your systems, your business know-how moves into the tool, and getting it out becomes a project.

This article comes from the AI Corner podcast episode where Thomas Spitz, CEO of AI Partners, hosted Quentin Amaudry, co-founder and CEO of Mendo, and from what we see on client engagements. It covers what actually ties you to one tool, where your know-how should live, and which open standards reduce the risk.

What does technological sovereignty mean for a company?

It is the ability to change AI tools or models without losing what your teams have built: their agents, their procedures and their connections to your systems. It is measured by the cost of leaving, not by the country the tool comes from.

The term is confusing because it is almost always used at the level of states. The European Commission frames Europe's tech sovereignty around semiconductors, cloud, AI and data. Quentin Amaudry draws the line in the episode: technological sovereignty, "not national, technological".

It is not the same as sovereign AI as hosting providers sell it, which is about where servers sit and which jurisdiction governs the data. Hosting a model in your own region settles the data question. It says nothing about where the instructions that make your agents work are stored, or whether you can take them elsewhere.

Why do AI agents make you more dependent on one tool than chat?

Because an agent encodes a business process inside the tool that runs it. A prompt can be rewritten elsewhere in five minutes; an agent that chains steps, business rules and several connectors has to be rebuilt, retested and signed off again by the people who use it.

Quentin Amaudry sees this with Mendo's customers, close to 200,000 users: once you start building agents and putting processes inside AI tools, he says, moving from one tool to another becomes much harder. His advice is blunt. Once the processes are in place, don't switch tools lightly.

The numbers point the same way. Menlo Ventures surveyed 150 technical decision-makers in mid-2025 and found that only 11% of teams had changed model providers in the past year, while 66% upgraded to a newer model from the same company. Switching is still technically easy, and in our view that inertia grows with every agent built inside the tool.

Should you standardise on one AI tool?

Yes if it helps you move fast, as long as what has value stays outside the tool. The risk is less about having a main tool than about storing know-how there that you could not take back.

The market moves too fast to bet on a winner. Menlo Ventures estimates OpenAI held 50% of enterprise LLM spend in 2023; by late 2025, Anthropic held about 40% and OpenAI 27%. For Thomas Spitz, whether you use one, the other, or one then the other matters less than building the habit, because a team that knows how to work with AI switches easily.

In large enterprises, the default is still Microsoft. Quentin Amaudry estimates 80% of the deployments he supports run on Copilot, because security and identity governance already live in the platform. The open question is whether being tied hand and foot to a single tool stays reasonable as token costs grow, a line item we break down in our article on agentic AI costs.

Where should your agents' business know-how live?

In a repository you control, in an open format that several tools can read. The AI tool then consumes your know-how instead of being the place where it is stored.

Quentin Amaudry gives an example in the episode: a Mendo customer keeps all its skills and sub-agents in a GitHub repository, which Mendo connects to directly. The format already exists. A skill is a folder containing a SKILL.md file with instructions, which Anthropic published as an open standard in December 2025. By October 2026, more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, had integrated it.

We apply the same principle with clients. At ISAGRI, product teams built a skills catalogue deployed across four code repositories, already picked up by other teams. At L'Oréal O+O, dozens of internally built skills are in daily use four months after the hackathon. In both cases the know-how belongs to the company.

Which open standards reduce your dependence on one tool?

Three standards cover most of it: Agent Skills for know-how, MCP for connections to your tools, and AGENTS.md for coding-agent instructions. All three now sit under the Agentic AI Foundation, part of the Linux Foundation.

The Model Context Protocol is the most mature. Anthropic donated it in December 2025 to the Agentic AI Foundation, co-founded with Block and OpenAI, noting it had already been adopted by ChatGPT, Cursor, Gemini, Microsoft Copilot and VS Code. OpenAI contributed AGENTS.md, which it says more than 60,000 open-source projects have adopted.

The model is the least worrying layer: the platforms themselves offer several, as Copilot shows. What is expensive to move is what you built on top.

How do you keep control without slowing adoption?

By setting a few rules from the first agents instead of waiting for a perfect architecture. Put in place before dozens of skills are circulating, they slow nobody down.

Any skill used by more than one person moves to the common repository with an owner, which also prevents agent sprawl. Connections to your tools go through MCP rather than proprietary integrations, because a standard connection plugs into another tool without being rebuilt. Every agent also has an owner, without whom nobody could migrate it: that is a prerequisite for AI agent governance, and often the job of an AI operator.

Reversibility also belongs in your contracts. For European operations, the EU Data Act, in application since 12 September 2025, sets rules for switching between data processing service providers and removes switching charges, including data egress charges, from 12 January 2027. Check with your legal team what applies to your AI tools, and require export of your agents and their configuration.

Watch the full episode

Thomas Spitz, CEO of AI Partners, hosts Quentin Amaudry, co-founder and CEO of Mendo, on driving AI adoption when the technology changes every week, the role of managers, the know-how teams hand over to agents, and technological sovereignty. The conversation is in French.

Frequently asked questions

What is the difference between sovereign AI and technological sovereignty?
Sovereign AI is about where models and data are hosted and which jurisdiction applies. A company's technological sovereignty is about whether it can change tools without losing its agents, skills and connections.

Is depending on one AI tool always a risk?
No, it is often a deliberate trade-off, because building on the platform that already holds your security and identities saves time. It becomes a risk when your business know-how exists only inside that tool.

Where do you start reducing dependence on one AI tool?
Inventory your shared agents and skills, where they are stored and who owns them. The most used ones are the first to move out of the tool's interface and into a repository you control.

Mapping what your teams have already built and deciding what should belong to you is the work of our AI First Organization engagement, and what we take to production with AI Factory. Book 30 minutes and we will look at where your know-how lives today.