AI Governance: From Experimentation to Impact in a Global Group (2027)
7 min read

AI Governance: From Experimentation to Impact in a Global Group (2027)

3,000 agents, a central hub, two governance bodies and an agenda owned by the business: Coty's experience.
AI governance: from experimentation to impact at Coty - AI Partners

In a large group, AI rarely comes in through a single door. Licences get handed out, teams build their own agents, vendors promise spectacular gains, and two years later nobody really knows what's running or who maintains it. That's when AI governance becomes a leadership issue.

In this episode of the AI Corner podcast, Thomas Spitz, CEO of AI Partners, talks with Nicolas Comestaz, Vice President of the Global Data & AI Center of Excellence at Coty, one of the world's leading beauty companies. His goal fits in one line: moving from the era of experimentation to the era of impact.

He explains how Coty listened before buying, let its teams experiment, built a hub-and-spoke organization, then took back control with an agenda owned by leadership.

Watch the episode on YouTube (in French)

Why does enterprise AI start out as a sellers' market?

Because supply arrives before demand. When he took the job, Nicolas Comestaz started by mapping the group's data ecosystem, from supply chain to finance and R&D. Then generative AI arrived with a lot of noise, and his message to the teams was clear: this is a sellers' market. He sees the same dynamic as adtech and martech in the 2010s, only more intense.

His rule: define the need, the purpose and the expected value first, then decide what to buy and what to build. His years at LVMH convinced him that data should be owned: bringing analytics back in-house there made leadership meetings with Google or Meta far better prepared. With AI the stakes go further, because you have to get inside business processes to find the value.

How do you roll out AI without losing control?

Give the tool to people who ask for it. At Coty, around 1,000 Copilot licences were first assigned on employee request. That phase surfaced use cases and, above all, people: curious users, and more technical ones able to build business applications, including on Azure AI Foundry. In parallel, a few group-level projects went straight to scale, such as content adaptation with Pencil.

A licence doesn't create usage on its own. Coty launched a program called Supercharge with AI: on-the-ground sessions in every market, a first generic level (what's possible, the mistakes to avoid, prompting), then a second one on the Copilot ecosystem. All of it repeated again and again.

Training comes with a global AI charter and policy: values, good practice, individual responsibility and the requirement to use the group's official tools. Thomas Spitz treats this as phase zero of any rollout, without which usage spills over into shadow AI.

What organization and which bodies should govern AI?

Coty chose a hub-and-spoke model. The central hub brings together four capabilities: the center of excellence, data governance, master data and a Digital & AI Factory. In each function, a Data & AI domain owner sets a strategic agenda and launches the data and AI products for that business. The center creates the framework and the capabilities, and business expertise puts them to work.

Two bodies keep it running. Every use case a function deploys goes through the AI Board: teams come to present it, and the board checks that it doesn't put the organization at risk. The Data & AI Circle brings the domain owners together every month to share what works and steer them toward the strategic agenda. Nicolas Comestaz sums up this first period in two words: controlled bottom-up.

When should you move from experimentation to top-down?

When initiatives multiply faster than you can maintain them. At Coty, more than 3,000 agents have been created, and sometimes ten of them do the same thing from one market to the next. The message to teams changes: experimenting is fine, but it isn't their main job. Coty's job is to make and sell beauty products, and AI remains an accelerator.

Hence the second phase, now under way: an agenda sponsored, or even led, by the head of each function, which allocates resources according to expected benefits and standardizes some ways of working. Thomas Spitz describes the same turning point at a client we recently worked with: very high individual usage, very low collective usage, and one employee who no longer knew who had given them the agent that had stopped working. You then need roles to maintain agents, like the AI operator, and proper AI agent governance.

Why does cutting the number of tools speed up adoption?

Because every new tool adds a journey and a change to support. For Nicolas Comestaz, the decisive advantage of the Microsoft ecosystem was meeting users where they already work, in Teams, with Copilot as the entry point. First seen as an office tool, less innovative than ChatGPT, Copilot proved its value through the ecosystem as a whole, up to Copilot Studio and Azure AI Foundry, under the security of the enterprise contract.

Thomas Spitz quotes a line that goes around: past fifty or so AI tools, you stop gaining productivity. And switching tools four times means four times the change management for a quarter of the adoption. The Content Factory we're building to generate visuals runs as a web app and directly inside Teams for that reason.

What profile should lead AI transformation in a large group?

A bilingual architect: someone who understands the business, knows enough technology to challenge, and can bring the organization along. Nicolas Comestaz puts the technical share of the role at around 30 to 40%, not to do the work, but to influence the business and coordinate technical teams. Like an architect, you coordinate a complex building site without knowing how to lay tiles or build the roof.

Thomas Spitz prefers the image of the translator, who explains a business need to technical teams and then makes what was built understandable to the business. Both land on a role that also owns the overall vision. Data and AI transformations often report to IT leaders, who are at the same time gaining a seat on the executive committee.

What's left to transform once AI is adopted?

People and processes. In Nicolas Comestaz's words, the transformation has moved from a technology transformation to a human and process one. That means redesigning the workforce, redeploying people to new tasks and adopting more rigorous ways of working, so automation rests on solid foundations.

Thomas Spitz shares the diagnosis: you can't keep stacking layers on an existing process, you have to rethink it. Often by building the new one alongside the old, then making the switch. That's the path to an agentic organization.

Structuring AI organization and governance is the work of our AI First Organization engagement. We train teams through AI Academy and put agents into production with AI Factory. Book a 30-minute call and we'll map your first three AI use cases.

The episode in 60 seconds

Key moments from the conversation with Coty, in short format.

Watch the short on YouTube

See more shorts on YouTube

Frequently asked questions

Who is Nicolas Comestaz?
Nicolas Comestaz is Vice President of the Global Data & AI Center of Excellence at Coty, a function created a little over two years ago to speed up the use of data and AI across the group's markets and functions. He previously led data and analytics transformations in media and digital marketing, including at LVMH.

What is AI governance in a company?
It's the set of rules, bodies and roles that decide which uses of AI are allowed, which come first and who maintains them over time. Compliance is part of it, but governance that stops at compliance doesn't tell you where to put your resources.

How do you set up an AI governance committee?
Start with one body that reviews every use case for security and compliance, like Coty's AI Board. Then add a monthly meeting with the AI leads in each business function, on the model of the Data & AI Circle, to align their projects with group priorities.

Who should own AI governance: IT or the business?
Both, with different roles: a central team sets the framework and shared capabilities, and the head of each function owns the use cases and decides the priorities. That's the hub-and-spoke model Coty chose.