A company adopts an assistant to help prepare its replies to customers. Its staff get used to it, then procedures move into the tool, including the exceptions only the most experienced colleagues knew about. One day, the supplier raises its prices or withdraws the model in use. The company then discovers what switching would cost: the subscription was only a small part of what it had committed, and some of its expertise now resides in the tool.
Nobody in that company decided to depend on that supplier. Each step seemed reasonable, whether to save time or avoid building something, and at each step another decision was possible. That is what we mean by choice.
Every organisation depends on technologies and expertise it does not produce itself, and most of these dependencies are reasonable. But it still needs to understand their consequences and know how to leave them behind. For us, sovereignty means choosing our dependencies and keeping the means to reconsider those choices.
Where value accumulates
In an AI project, an organisation contributes its data, its people’s time and its knowledge of the work. Without that knowledge, nobody could tell a correct answer from one that merely looks correct.
We design our projects so that the organisation retains something when the project ends, particularly written business rules and criteria for assessing answers. These gains must remain useful if the model or supplier changes. When we compare solutions, we therefore also calculate what it would cost to adapt or replace them.
The whole chain also needs examining. A service hosted in France may rely on a model whose access is still controlled by a supplier outside Europe, and data often travels further than expected, to the supplier or its subcontractors. There is also the question of how to keep working if the tool becomes unavailable.
For a public administration or a defence unit, this is already a practical question. In June 2026, an American supplier suspended access to its most advanced models for almost three weeks to comply with export controls. The law also requires public authorities to be able to disclose to a person the rules of an algorithmic process on which a decision concerning them was based. We incorporate these constraints into the architecture from the outset, alongside performance.
Keeping the ability to judge
The least visible dependence concerns skills. We call it cognitive dependence: an organisation delegates tasks and gradually loses the means to judge their results.
An early sign is often a legal professional who can no longer challenge the assistant they use. Then reviewing becomes a formality, and new employees learn to query the tool without learning what they would need to check it.
Dashboards show none of this: when the correction rate for an assistant falls to zero, there is no way to know whether the tool has become excellent or whether nobody reviews it any more. Having a human approve the result changes little if that person lacks the time, skill or authority to challenge the answer.
The tool can also help maintain that ability. An assistant that cites its sources and flags what it is unsure of requires its user to form their own judgement about the case. That is how we design its use.
Putting results to the test
Research in physics left us with a habit: before starting work, state what we are trying to demonstrate, how we will measure it and what would change our minds. We then test approaches on the client’s real cases, starting by looking for the ones where they fail.
When we consider an agent, meaning software that uses an AI model to carry out a sequence of actions, we first compare it with simple automation or a statistical model. If it does not do better, we rule it out. We sometimes also conclude that a project does not justify using AI at all, and we say so before any commitment.
For a tool that handles requests, we measure the time needed to complete a case correctly, including review and corrections. The choice between existing software, adapting it and building something bespoke follows the same approach, based on the full cost, including operations.
Time saved only has value if it serves a purpose, such as clearing a backlog or shortening response times. When a project’s effects can be measured and attributed, we propose making part of our remuneration depend on the results agreed together. We then share the risk and, like the client, have an interest in that time being put to good use.
Passing on control
All of this applies to us too. From the outset, the contract specifies what belongs to the client, what they can recover and under what conditions another team can take over the project. Documentation and training are part of the project, so that teams know the system’s limits and can judge what it produces.
When an internal AI team is useful to get started, we design it to pass its skills on to the business teams in about three years, with the timing depending on the organisation’s starting point. It has succeeded if, by then, those departments can manage without it. Some functions will remain with specialists, such as model operations or security, but judging a result and deciding what to do with it belongs to the business teams, who bear that responsibility.
Our work is to make these choices visible before they are made by default.
