Enterprise AI agents that lack an understanding of how a business actually operates are making flawed decisions at scale. Devavrat Shah, chief scientist of enterprise AI at Celonis, argues that the missing ingredient is operational context, and that without it even the most powerful models stumble inside complex workflows.

What You Need to Know

Enterprise AI has moved past the pilot stage, and leadership teams now demand proof that scaling will affect the bottom line. Devavrat Shah of Celonis argues that AI agents need a contextual layer connecting them to existing business processes and data before they can make reliable decisions. Without that layer, agents risk overlooking constraints, making costly mistakes or burning excessive computing resources. Shah recommends starting with one specific business decision, testing AI on its hardest steps and measuring gains before scaling further.

The Gap Between General Models and Specialized Workflows

Shah, who has taught AI at MIT for more than two decades, frames the problem as a mismatch between generalized principles and specialized scenarios. AI agents can support individual tasks within a larger decision-making process, he says, but they need to understand the role each task plays and the business conditions surrounding it.

"You can't just do that in a vacuum," Shah said during an episode of the Business Lab podcast, produced by MIT Technology Review in partnership with Celonis. Without the right operational context, agents risk overlooking important constraints, making costly mistakes or consuming excessive computing resources as they try to work out what they need to know.

His answer is what he calls decision intelligence: connecting AI to the processes and data enterprises already use. A contextual layer can help agents understand workflows, operate within defined guardrails and produce recommendations that people can review and approve.

Why This Matters

The stakes extend well beyond one vendor's product pitch. Companies that deploy agents without process context will likely see pilot programs stall, and the resulting skepticism could slow AI adoption across entire industries. Business leaders who demand measurable returns will judge the technology on decisions made, not tasks automated.

Employees stand to gain or lose the most. Automating repetitive manual work and coordinating tasks across teams could free people for more complex strategic decisions. Realizing that benefit, however, will require investment in workforce education so employees can interpret AI outputs and their implications instead of simply handing decisions over to agents.

Competitors in the process-mining and enterprise software space are watching closely. Celonis built its business on mapping how work actually flows through organizations, and Shah's argument positions that process data as the foundation for credible AI. Rivals selling standalone agent platforms will face harder questions about where their context comes from.

A Practical Playbook for Scaling AI Agents

Shah advises organizations to resist the temptation to deploy AI broadly from the start. His recommended sequence is deliberately narrow and measurable.

  • Pick one decision: Start with a specific business decision rather than an enterprise-wide rollout.
  • Find the hard steps: Identify the most difficult parts of that decision process.
  • Test the model: Check whether AI can genuinely improve those steps.
  • Measure and scale: Track gains in speed and value before expanding the application.

The logic is that context is easier to supply and verify at small scale. A single decision gives teams a controlled environment to see whether an agent respects constraints and produces recommendations people trust.

From Routine Tasks to Enterprise-Wide Planning

Shah sees the opportunity extending from routine operational work to enterprise-wide planning in finance, workforce management, marketing and sales. With the right context, AI could help organizations evaluate options and make consequential decisions more quickly.

Resource planning is one example he cites. "Now, maybe [resource planning] should not take that much time, and people are actually making the right choices quickly," Shah said. The shift he describes is not about replacing judgment but about compressing the time between question and decision.

That vision depends on enterprises treating context as infrastructure rather than an afterthought. Organizations that map their processes and connect them to AI systems will be positioned to scale agents responsibly. Those that skip that step may find their models fluent in general knowledge but blind to the realities of their own operations.