"Scaling up and rolling out AI is a human issue, not a technological one"

Published : 16/07/2026

Olivier Serfaty, Director of Data, AI and Innovation at Inetum France, heads up the group’s Data & AI Factory, which was launched in early 2026 following more than 120 pilot projects, 30 of which were rolled out at scale in just over two years. He shares his insights on the real keys to scaling up.

A modern office setting where a group is gathered around a table with laptops and documents while one individual stands and gestures during a meeting, illustrating a workplace collaboration scene.

Why do some AI projects, despite being promising, never make it past the testing stage?

Olivier Serfaty: At Inetum, we have formalised our approach in what we call Coborg ™, which stands for ‘Cognitive Brain of your Organisation’. It is a strategic framework based on five pillars that structures the process of scaling up. The first pillar is governance. This isn’t just a technological issue; it’s first and foremost an organisational one. For a project to scale up, the right stakeholders must be brought to the table from the outset: the IT department, which provides the data and ensures security; a Chief AI Officer, who oversees technological decisions; HR, because the deployed agents now carry out tasks in addition to those performed by humans; and finally the business unit, because it is the business unit that has the need and must explain it. When the business unit works in isolation, or when the IT department does the same, the use case is not deployed on a large scale.

Once you’ve brought the right people together, where do you start?

Olivier Serfaty: By selecting the relevant use cases. At Inetum, we have developed a tool called the Entropy Tool. It is an AI system that breaks down a process node by node and assesses at each stage whether AI can automate a task, whether the data is available and reliable, or whether that stage can only be carried out by a human. This produces a ‘score’ that clearly indicates whether there is a real benefit to developing an AI solution or not. This prevents us from deploying a tool that would not deliver any productivity gains.

In practical terms, what constitutes “ready” data for an AI project?

Olivier Serfaty: There are three aspects. Firstly, traceability: we need to know where the data comes from, who has modified it, and whether it is reliable and up to date. Secondly, security: the models used must be private; the data cannot be freely available on the internet. And sovereignty: many clients have hosting restrictions, either for regulatory reasons or because they do not want their data to be transferred to the United States. In Europe, we also have the AI Act, which imposes very strict traceability requirements in certain regulated sectors (notably healthcare, finance and recruitment). We must therefore be able to explain why a particular result was obtained, based on which data and using which model.

What’s changing, then, isn’t that humans are disappearing—it’s that their value is shifting from the accumulation of knowledge to judgment. A criminal lawyer used to spend three days locked away in their office poring over case law, but tomorrow, AI will provide them with a summary of the 56 key decisions they need to keep in mind. They can devote that saved time to thinking about the substance of the case. AI automates a large part of their work, but they retain a critical eye, and it is always they who make the final decision. Yet to decide is to choose: to be able to explain what is being ruled out as well as what is being retained, in order to refine a line of argument in a courtroom case, for example. This is a matter of free will, and the machine does not possess it.

How do you assess the state of a client’s data before getting started?

Olivier Serfaty: Before generative AI, this assessment used to take several weeks. We have created a tool that uses generative AI to analyse the traceability of data within the company’s systems: SAP, Salesforce… The model establishes links between all the data, identifies its sources and produces a comprehensive map. We obtain an overview in a matter of hours rather than several weeks’ work. Some clients contact us solely for this data assessment, even before they know what use case they will implement next.

The ultimate aim is for a business unit to be able to query its data using natural language, without writing a single line of SQL. For example, by asking, ‘What was my turnover last year in a particular region, for a specific product, and how has it changed over the last five years?’ and receiving a result straight away. We’re not quite there yet, but that’s the direction we’re heading in.

How can we ensure the reliability of the results once the solution has been rolled out?

Olivier Serfaty: This is a critical point. A law firm that analyses contracts using AI cannot afford to have a single word misinterpreted by an LLM: a single word can change an entire clause. We use an approach known as ‘LLM as a Judge’: we submit the same question to three or four models simultaneously, then an additional model analyses their responses, merges the converging answers, flags any contradictions to a human, and detects any potential hallucinations by cross-checking. The aim is to achieve a relevance rate of over 99 per cent.

Once the tool has been rolled out, how do the business units adapt in practice?

Olivier Serfaty: This is Coborg’s final pillar, and perhaps the most important one: change management. Take a sales adviser, for example. When a customer calls, the AI agent automatically displays all the relevant information on their screen. It suggests responses. Once the call is over, emails are sent, an order is placed on SAP, and the CRM is updated automatically… This sales representative is no longer doing the same job. This represents a productivity gain of at least 30 per cent. The question then becomes: what should be done with this freed-up time? We are no longer in the awareness-raising phase of two years ago, when we had to explain what AI was. People have fully understood its benefits. Now, we need to support them through the practical transformation of their roles, their processes and their day-to-day work.

Request a contact