Enterprise AI systems must meet a fundamentally different standard than consumer chatbots: they need to be explainable, interpretable, and reliably repeatable because the stakes are higher and the accountability is clearer. At HumanX Amsterdam, Ambica Rajagopal, group chief data and AI officer at Michelin, emphasized that companies deploying AI internally cannot tolerate the error rates that consumer products like ChatGPT launched with, because the consequences ripple through operations and decision-making in ways that affect employees, customers, and the bottom line.

Why Does Interpretability Matter More in Enterprise Settings?

The distinction between consumer AI and enterprise AI hinges on accountability and trust. Rajagopal noted that releasing a product with ChatGPT's initial error rate inside Michelin would have been catastrophic. "My phone would not have stopped ringing," she said, underscoring the operational reality that enterprise deployments carry expectations of accuracy, reliability, and measurable return on investment that consumer products simply do not face.

Rajagopal

This gap reflects a deeper truth about how AI is used in different contexts. Consumer copilots assist individuals with tasks they could theoretically do themselves. Enterprise AI systems, by contrast, often make decisions or perform work autonomously, with minimal human oversight at every step. That shift in autonomy demands a corresponding shift in rigor. When an AI system is doing the work rather than being checked by a human at every step, the bar for explainability and interpretability rises sharply.

"At the end of the day, an AI model is a function. Explainability, interpretability and repeatable behaviour are therefore things enterprises are entitled to demand. It just takes research and effort," said Ambica Rajagopal.

Ambica Rajagopal, Group Chief Data and AI Officer at Michelin

Rajagopal holds a PhD and brings a technical perspective to the problem. Her framing is elegant: if an AI model is fundamentally a function, then enterprises should expect the same level of predictability and transparency from it that they would from any other critical business function. That expectation is not unreasonable; it simply requires investment in mechanistic interpretability and explainability research.

How Are Enterprises Building Interpretable AI Systems?

The path forward for enterprise AI involves several key practices that prioritize transparency and reliability:

• Custom Model Development: Rather than relying on off-the-shelf copilots, Michelin chose to build custom models tailored to its own value chains and processes. This approach gives the company direct control over model behavior and the ability to understand how decisions are made within its specific context.
• Runtime Reliability Techniques: When AI systems operate autonomously, new technical approaches are needed to ensure reliability at runtime. These go beyond traditional testing and require continuous monitoring and validation of model outputs in production environments.
• Clear Ownership and Accountability: Enterprises must establish who owns the AI system and who is responsible for its behavior. This clarity enables accountability and ensures that explainability requirements are not treated as optional add-ons but as core design principles.
• Process-Level Redesign: Rather than making individual workers slightly faster with copilots, enterprises are reimagining entire workflows around AI capabilities. This requires understanding not just how the model works, but how it fits into and transforms the broader organizational process.

Michelin's approach illustrates this in practice. The company has more than 30,000 people using its internal generative AI platform, but Rajagopal estimates that accounts for only about 20 percent of the value AI delivers to the organization. The remaining 80 percent comes from custom models embedded in processes that no other company has: predictive models in manufacturing, generative models for tire design, and models supporting complex forecasting operations.

Over three years, AI has returned more than $200 million in value to Michelin's business, growing around 30 percent annually. That scale of return is only possible because the company invested in understanding and controlling its AI systems at a deep level, not because it deployed more copilots.

What's the Difference Between Copilots and Digital Workers?

The distinction between these two categories of AI systems shapes how interpretability requirements differ. Darko Matovski, co-founder and CEO of causaLens, drew a clear line: "Copilots make an individual more productive. Digital workers do the work".

A copilot is an assistant that augments human capability. A digital worker is an autonomous system that replaces or compresses work that would otherwise require multiple teams and weeks of effort. The economic case for digital workers is stronger, but the interpretability demands are also higher. An assistant might make someone 20 to 30 percent more productive, but token spend is unpredictable; a single prompt can cost anywhere from $100 to $10,000, making return on investment hard to forecast and capped by what people do with the time they save.

Digital workers, by contrast, can compress entire workflows. Forecasting demand for individual products, or SKUs, can involve around 20 teams and a stack of software. Rather than making each person slightly faster, a digital worker can cut across all of it, eliminating redundancy and accelerating decision-making. But that level of autonomy requires that the system be interpretable and trustworthy.

Why Is Leadership the Real Barrier to Scaling Enterprise AI?

When asked what actually stops enterprises from scaling AI, Matovski gave a one-word answer: leadership. When causaLens is given a mandate to redesign a process, it succeeds. When it is limited in what it can change, it does not. The conversation has moved up the organizational hierarchy, from heads of data science to CIOs, then CFOs, and now to chief executives who say they rolled out everything their vendors recommended and still cannot see the impact.

This pattern suggests that the barrier to scaling enterprise AI is not technical. The tools exist. The challenge is organizational: companies must be willing to rethink processes from the ground up, to invest in custom models rather than buying commodities, and to demand that their AI systems be explainable and interpretable. That requires leadership commitment and a willingness to treat AI as a strategic capability, not a tactical tool.

The HumanX Amsterdam event revealed that early-stage startups are beginning to address this gap. Companies like LangWatch, which builds infrastructure for testing and evaluating AI agents, and 8wave, which runs an AI governance platform that gives companies visibility of their AI systems and audits them against regulation, are creating tools that make interpretability and explainability more accessible to enterprises.

As enterprise AI adoption reaches record highs, the conversation is shifting from "Can we deploy AI?" to "Can we understand and trust the AI we deploy?" That shift puts interpretability and explainability at the center of enterprise AI strategy, not at the margins.

" }