- Sophie Dionnet, Dataiku’s SVP of Product and Business Solutions, identified people, orchestration and governance as the three critical ingredients for enterprise AI value at the Dataiku Summit in Singapore in July 2026.
- Dionnet argues governance functions as a scaling mechanism, not a brake on innovation, enterprises that treat it as a compliance overhead pay higher remediation costs later, particularly in regulated industries.
- A Roche patent lawyer encoding his own expertise into a working AI agent system illustrates Dionnet’s core point: domain knowledge, not model capability, is usually the binding constraint on enterprise AI deployment.
Most enterprises aren’t losing the AI race on model capability, they’re losing it on change management, governance and the wrong people in the room. That was the central argument Sophie Dionnet, SVP of Product and Business Solutions at Dataiku made at the company’s Summit in Singapore in July 2026, where she laid out a three-part framework for moving AI from pilot to P&L impact.
People First
Dionnet places people at the foundation, but her definition goes beyond hiring data scientists. The profile she’s looking for combines technical depth with business domain expertise, a pairing that’s rarer and harder to develop than either skill alone. Her sharpest example: a patent lawyer at Roche who encoded his own professional knowledge into a working system of AI agents. The technology didn’t drive that outcome; the domain expertise did.
Change management, she argues, is where most implementations actually stall. Even after a board-level decision to adopt AI, full deployment can take years. Companies consistently underinvest in the human side of AI adoption upskilling business professionals, building cross-functional teams and creating the conditions for sustained use, while continuing to spend on new models and tools. That imbalance is where value leaks.
Orchestration Across the Stack
The second ingredient is orchestration: the capacity to connect disparate technologies, legacy systems, data infrastructure, LLMs, agents, into something that operates reliably at scale. Dataiku positions its platform as that orchestration and governance layer, unifying people, data assets and AI tooling under a single governed environment.
Dionnet’s concern here is “agent sprawl”, the proliferation of AI agents across an enterprise without consistent monitoring or control. The technical stitching matters, but she flags a failure pattern where IT teams over-index on integration architecture while the business process implications go unexamined. Effective orchestration, in her framing, means AI that is deployed, monitored and maintained in production, not just built. For teams working through agent orchestration at scale the governance layer is increasingly where the hard decisions sit.
Governance as Infrastructure
The counterintuitive argument Dionnet makes most forcefully is about controls. Governance, covering data quality, model explainability, fairness and oversight, is routinely treated as overhead or a compliance checkbox. She inverts that: in regulated industries like financial services, the cost of fixing poorly governed AI after the fact consistently exceeds the cost of building governance in from the start.
The organisational tension she describes is familiar. Fully decentralised AI experimentation produces conflicting outputs, diluted efforts and accountability gaps. Excessive centralisation through IT slows deployment and removes the business context that makes AI useful. Dionnet’s recommended middle path is embedding governance throughout the AI lifecycle rather than bolting it on at the end, a structure that allows experimentation without compounding risk as deployments scale. Multi-agent deployments at enterprise scale are making this balance harder to maintain without purpose-built controls.
The Model Isn’t the Bottleneck
Dionnet characterises the post-2022 LLM wave as a gold rush, organisations attracted to new model capabilities before they have the internal machinery to deploy them. The gap between deciding to adopt AI and realising value from it is, in most cases, not a model problem. The latest frontier model is rarely what’s blocking an enterprise from extracting value; organisational capacity for change is.
That reframes where leadership attention belongs. Continuously evaluating new models while the deployment infrastructure remains underdeveloped is a common pattern. Building the internal capability to operationalise existing AI reliably tends to compound faster than model-chasing does.
From Pilots to P&L
On moving from experiment to measurable financial impact, Dionnet is pragmatic. Not every AI initiative will produce multi-million dollar returns, and she doesn’t think that framing serves enterprises well. The more useful capability is a systematic way to measure and act on varied levels of impact, incremental operational improvements alongside larger transformational ones.
The companies that scale AI successfully, in her account, typically start with a small number of clear wins that build internal trust. Those early pilots then need to reach production and demonstrate real operational or financial improvement before broader investment is justified. Without that proof chain, the default outcome is a portfolio of inconclusive experiments.
Broadening Who Builds AI
Generative AI’s lower technical entry point is changing who participates in AI development. Dionnet sees this as an opportunity to bring business professionals with deep domain expertise, people who previously couldn’t engage with the technical stack, into the build process. The Roche example is the clearest illustration: a domain expert who could not have contributed to AI development five years ago is now building agents.
Dataiku’s platform strategy is built around this premise, visual tooling for business teams, coding environments for data scientists, all within a governed architecture. The framing Dionnet uses for where this leads is direct: as organisations work alongside AI agents, the differentiating human capabilities shift toward judgment, collaboration and governance rather than raw technical execution.
