Florian Douetteau

The through-line

Across more than a decade of arguing the same case, Douetteau's position has stayed fixed: enterprise AI only pays off when the people who own the business process build the thing themselves. Dataiku was founded in 2013 on what its own telling calls a contrarian thesis, that enterprise AI transformation has to come from business operators rather than centralized data science teams, with the product acting as the translation layer between them [6]. The same idea was framed earlier as "Everyday AI," a platform where data experts and domain experts work together so that AI becomes mainstream through the collective effort of the whole company instead of a technical elite [4]. What has changed is the object being built. The collaboration used to produce models and analytics inside daily operations [4]; it now produces agents, and the platform's job has extended into cataloging data, orchestrating fast-moving technologies, and governing what gets deployed [1][6].

The urgency has also sharpened. Where the early argument was about accessibility and impact, the current one is competitive survival: "the risk for many organization is to be left behind and the reason why they need to focus on their core workflows, their core processes and transform them rapidly with AI is because if they don't do that, they will be displaced by others very rapidly" [1].

On why use cases are the easy part

Douetteau's most direct claim about the current market is that finding value is no longer the bottleneck. Large enterprises "struggle in terms of understanding oh they should scale AI and it's not just about like figure out the use cases it's become the easy part like everyone is finding value in AI" [1]. The difficulty sits in three dimensions that each carry their own failure mode: "you've got uh the people angle the orchestration angle and the governance angle and uh AI can fail across all each of those dimension like you don't actually have the right people to help you build AI you don't know to orchestrate and combine all of those fastmoving AI technologies and also do you govern AI or do to manage a risk" [1]. The people problem is the oldest of the three in his thinking, going back to the observation that there is a persistent shortage of people with the core technical skills to get things done, which is exactly why analytics has to be co-designed with the people who understand the business process rather than handed off to specialists [5]. Governance is the dimension Dataiku has increasingly built its position around, serving as the governance layer for its enterprise customer base [6].

On agents where the money is

The agent use cases Douetteau cares about are unglamorous by design. Dataiku focuses "on agents that are about those back office and operation function where the money where the efficiency really is" [1]: agentic optimization of supply, production operations in manufacturing so that assets are up and running on time, and credit risk operations in banks, where the aim is to "cut through the red tape and optimize lots of reporting processes" [1]. The common shape is "all of those uh lengthy heavy processes very human intensive where you have to manage all of the data and you spend lots of money and resources making that happen" [1]. The build pattern follows the founding thesis: customers such as Novartis, BNP and Standard Chartered use the platform to "enable their people to actually build agents by themselves," connecting their data, the available LLMs and their own business expertise so that "their existing business process" becomes an agent [1]. Standard Chartered features in his earlier accounts too, in the context of automating manual, in-flight operational work [5].

On managing agents at scale

The newer products, a data cataloging system and agent management, follow from a problem his customers hit once adoption works. "Many of our customers already have dozens to hundreds of agents and their question is are they doing the job? Are they actually working? Do they actually derive business value?" [1]. Agent management exists to observe and manage agents at scale, and he frames the coming management challenge in human terms: organizations are "today uh struggling as an organization sometimes to manage people. Tomorrow it will be about managing agents" [1].

On the automation cursor and humans in the loop

Long before agents, Douetteau was arguing that automation is a dial rather than a switch. He describes a spectrum running from decision support, where the system offers suggestions to a human decision maker, through partial automation with a human in the loop and human supervision, to fully automated processes where the work reduces to monitoring the model [5]. Where an organization sets that cursor is industry- and process-specific: routine card payments can be automated at high volume, while other situations demand a person, and each business has to decide for itself where the cursor sits given the implications for employment, for the culture of the company, and for the customer relationship [5]. He calls the resulting practice a new discipline, automation design, which requires a conversation between the analytics side and the heads of the business that in his view is not yet mature in most enterprises [5].

He is also wary of automating away the wrong thing. Using an extended comparison with the circus, he notes that a business runs on human heuristics built up through experience, and that automating them without first understanding them destroys the creativity and the people that made the process work in the first place; the goal is to understand which heuristics matter before replacing them [5].

On explainability as a business problem

Explainability, in his account, is not a checkbox that gets ticked once a technique is available. The technology to explain a classification model exists, but the harder question is where explanation enters the business process and who it is for [5]. He distinguishes the back-office use, where the modeling team checks its own work, from the deployed use, where an explanation has to reach an operator inside a business application such as a CRM and be intelligible in that context [5]. The integration problem is the real one: pushing a score into a business system is easy, but delivering the reasoning alongside it, in a form the person acting on it can use, is considerably more complex [5].

On one platform and the lifecycle tangle

Dataiku's architecture reflects a specific frustration: real organizations run several overlapping lifecycles at once, a data lifecycle, a model lifecycle with its build and operate and monitoring stages, and a project lifecycle with its own stakeholders, and these do not line up neatly the way tidy diagrams suggest [5]. His answer is a single platform that gives a consistent view across data, model and project rather than a set of tools that each cover one slice and leave the coordination to a project manager [5]. The 2026 refresh continues that logic, putting a platform for AI success into the market on the strength of what the largest enterprises are already doing with it [1].

On building an enterprise software company from outside Silicon Valley

Douetteau reflects on the difference in vantage point between building data software in Europe and building it in Silicon Valley. In the Valley, proximity to the same competitors and the same category conversations pulls companies toward de facto competition and toward positioning inside a narrow niche while trying to claim a grand category [5]. Building elsewhere, and serving big enterprises rather than startups from the start, meant working with customers who already had legacy systems, existing reporting and decades of accumulated infrastructure, and layering advanced technique onto that reality instead of assuming a clean slate [5].

On the CIO's make-or-break moment

His guidance for CIOs starts from their squeezed position: boards want AI to happen while worrying about AI risk, and business units want to buy and ship without necessarily understanding the technologies, the risks or the constraints [1]. That makes this "a make or break type of moment for them," where they either lead the transformation and "become more than a CIO" or get replaced by AI-first CIOs who capture the moment [1]. He draws the parallel explicitly: twenty years ago the advent of cloud produced a new generation of CIOs defined by having moved their company to the cloud, and the same generational sorting is now happening around AI [1].

Takeaways

• Identifying AI use cases is no longer the hard part; scaling fails on the people, orchestration and governance dimensions instead [1].
• Enterprise AI should be built by business operators who own the process, not by a centralized data science team, with the platform acting as the translation layer between them [4][6].
• The highest-value agents sit in back office and operations, supply, manufacturing uptime, credit risk and reporting, where lengthy human-intensive processes consume money and resources [1].
• Once customers run dozens to hundreds of agents, the question shifts from building to observing and managing them, and to proving they derive business value [1].
• Automation is a cursor, not a switch, running from decision support through human-in-the-loop to full automation, and every industry has to place it deliberately given the effects on employment, culture and customers [5].
• Automating human heuristics without first understanding them risks destroying the creativity that made the process work [5].
• Explainability only counts when it reaches the operator inside the business application, which is an integration problem rather than a modeling checkbox [5].
• CIOs face the same generational sorting that cloud produced twenty years ago: lead the AI transformation or be replaced by AI-first successors [1].