.png)
What's in this article :
1. Why the order in which you deploy AI changes everything?
2. What are the concrete strategies to get your teams to adopt AI?
3. How to bridge the gap between business teams and data teams?
4. What results can you expect when teams are involved from the start?
5. FAQ
6. Conclusion
AI deployment order refers to the sequence in which a company introduces tools, trains its teams, and defines its use cases. The majority of companies deploy first and train afterwards. This is precisely the order that explains most failures.
Thierry Champerou, CDO and Director of Data and AI at MAIF, frames the problem clearly: "Many companies have used AI to completely eliminate customer service. The promise was: AI will answer you in real time. In reality, we end up with no customer service at all. It is unbearable."
This is not a theoretical observation. It directly impacts customer relationships. A chatbot deployed without involving the teams who handle real customer requests produces correct answers only half the time.
The distinction Thierry Champerou draws is fundamental: automation versus assistance. Automating means removing the human from the process. Assisting means increasing their ability to make the right decision. "The zero-risk threshold in generative AI does not exist today. This is why I strongly prefer the term assistance over automation."
In sectors where a wrong answer has direct contractual consequences, generative AI cannot bear that responsibility alone. The employee must retain their professional expertise to correct, adjust, and decide. This principle applies beyond insurance to any sector where the quality of the response commits the organisation.
AI adoption in enterprise rests on a counter-intuitive principle: the earlier you involve teams, the less resistance you encounter. Not because it feels inclusive, but because it is technically the only way to obtain use cases that actually work in the reality of each role.
Before selecting a solution, run a collective understanding phase. The objective: every employee understands what AI can do, what it cannot do, and where it produces errors. A team that does not understand the hallucination risk cannot exercise the critical judgment required to use these tools correctly.
MAIF applied this principle through its Employees' Convention in 2024. 30 employees selected by representative draw from across the entire group, representing all departments and age groups, attended 2 days of training delivered by external specialists before working on recommendations. "We dedicated the first two days entirely to training and acculturation. And I was genuinely amazed by the quality of the work produced."
Thierry Champerou is direct on this point: "Building AI solutions means working with the real people. The real people in the field." A functional specification written by a project management office at headquarters does not capture the reality of a customer advisor handling 80 calls a day.
This proximity is not a methodological detail. It determines whether the solution will be used or bypassed. A tool that does not solve a real operational problem will never be adopted, regardless of its technical sophistication.
In any large organisation, there are early adopters who immediately embrace AI tools, and a more cautious or reluctant majority. The mistake is to design the deployment solely for the former. "You need to defuse things, you need to accompany people. The first priority is to train, to build familiarity, and to encourage a controlled use."
MAIF deployed Copilot Web to all its employees, not primarily for immediate productivity gains, but as a progressive acculturation tool. The primary objective: give everyone access, let them experiment, and let them develop their own relationship with the tool before formal use cases are imposed.
The gap between business teams and data teams is the most underestimated obstacle in an AI deployment. It is not a technical problem. It is a language and culture problem.
"It is a cultural shock. We do not speak the same language. We do not even realise how much we use jargon in the world of AI." This observation from Thierry Champerou touches a point that many management teams overlook when launching their projects.
Data scientists previously owned their models: controlled training data, explainable results, transparent logic. Generative AI has disrupted this balance. LLMs are trained by major American actors on datasets that no internal team controls. "In itself, this is a Copernican revolution for our data scientist colleagues."
Three concrete actions to reduce this gap :
Results are measurable across two dimensions: the quality of what teams produce when they are genuinely consulted, and the durability of the AI systems deployed.
MAIF's Employees' Convention produced 40 recommendations in 4 days. 38 were retained by the executive committee, with one central commitment: despite the emergence of AI, no position would be eliminated. "Management had committed to listening and responding to these recommendations. And I was truly amazed. Questions that management itself would not necessarily have thought of."
This result illustrates something that is easy to overlook: the professional expertise of an advisor or a manager is not in the algorithms. It lives in the people who practise that role every day. Collective intelligence consistently surfaces use cases that top-down planning misses.
Thierry Champerou runs more than 50 AI services in production at MAIF, built on a data infrastructure developed progressively over nearly 10 years. That depth of data maturity is not an accident. It is the direct result of building incrementally, with teams who understood the systems they were shaping.
The lesson is practical: organisations that deploy fast without team involvement often reach production quickly and stall just as fast. The AI systems that last are those maintained in what Champerou calls "intelligent operating conditions": continuously monitored, adjusted as external models evolve, and kept aligned with the operational reality of the people using them.
This kind of durability cannot be bought off the shelf. It is built through the habits, the vocabulary, and the ownership that only come from genuine team involvement from the start.
Watch the full podcast episode
Watch the full AI Corner episode on YouTube
Resistance almost always comes from a lack of understanding, not a refusal in principle. Employees who have participated in an acculturation session and been consulted on use cases become adoption advocates. Those placed in front of a tool without explanation develop distrust and workarounds.
No. Start with a representative group covering different roles, not just tech-savvy volunteers. The goal is to produce use cases grounded in the reality of multiple departments, then spread the results. A company-wide deployment without a pilot phase produces high costs and diffuse results.
By going into the field with the teams who execute the processes daily. The best use cases emerge from direct observation, not management meetings. A use case that solves a real operational problem will be adopted naturally. One that solves a theoretical problem will not.
Automation removes the human from the process. Assistance increases their ability to make the right decision. In contexts where a wrong answer commits the organisation, assistance is the only viable model. The employee retains their professional expertise; AI allows them to go faster and further.
First results are visible within weeks when use cases are well defined with operational teams. Duration depends on existing data maturity, not solely on the tool chosen. MAIF's experience, nearly 10 years of incremental infrastructure and 50+ services in production, shows that the deepest results compound over time.
Involving your teams in the AI strategy before any deployment is not an organisational constraint. It is the condition that determines whether the tools deployed will produce real impact or remain unused. The winning sequence is always the same: collective acculturation first, use case identification from the field, then progressive deployment with continuous monitoring. For organisations looking to build this approach with structure and speed, AI Partners helps enterprises turn scattered AI initiatives into a structured roadmap, from maturity diagnosis to the deployment of first use cases, through its AI First Organization programme.
