Celonis's Devavrat Shah: AI agents lack context, underperform

Shah's comments come as enterprise AI moves from experimentation to everyday operations, a transition he describes as being at a crossroads after years of demos and pilots. His argument is that the technology's biggest impact will come in decision-making, where AI agents can support individual tasks within larger processes rather than replace whole decisions. He frames this as bringing “decision intelligence” to enterprises, an approach that depends on connecting AI to the processes and data companies already use.

Celonis's background is in process intelligence, helping organizations adhere to ideal processes and re-engineer them. Shah describes extending that to decision intelligence by having agents take on atomic steps that experts previously handled, with a context layer designed to abstract the systems below and guardrail the agents above. He notes that connecting to existing systems of record and workflow tools requires classical information security review before a vendor's software can be trusted inside an organization.

Looking ahead, Shah expects the shift to free people from repetitive manual work and coordination tasks so they can focus on strategic decisions, while cautioning that workforce education must keep pace. He points to enterprise-wide planning in finance, workforce management, marketing, and sales as areas where faster, better-informed decisions could matter most, suggesting annual planning cycles may no longer need to take a full year.