Enterprise AI's Missing Layer: Operational Context
Powerful models still fail inside real businesses. A context layer that maps actual workflows may be what finally bridges the gap.
By The energy-climate Desk
Enterprise AI can generate recommendations, automate tasks, and help organizations make decisions faster — but without understanding how a business actually operates, even powerful models miss the constraints that determine whether a decision is the right one. [1] That gap is now the central problem for companies trying to move AI from pilots into daily operations. [1]
Devavrat Shah, chief scientist of enterprise AI at Celonis, frames the problem around what he calls "decision intelligence." AI agents can support individual tasks within larger decision-making processes, he argues, 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," he says. [1]
Without that grounding, the risks compound quickly. An agent operating without operational context will either burn through excessive compute trying to reconstruct what it needs to know, or — more dangerously — miss a critical constraint and act on it. [1] Shah draws a sharp distinction between a bad restaurant recommendation and a flawed financial forecast: one is a minor inconvenience, the other can contribute to a company's collapse. [1]
[2] It combines end-to-end process data, business knowledge, and intelligence so that people and agents have "the operational clarity they need to reason correctly, decide sensibly, and act reliably." [2] The architecture sits between a system-of-record layer below — workflow tools and event logs — and an agentic layer above, where AI agents or human experts consume the output. [1]
[1] Fujitsu uses Celonis to gain full transparency into inventory levels and AI-enhanced recommendations about when to stock-up, re-allocate, and refrain from buying materials; Uniper is transforming how work gets done across the enterprise with Celonis Process Intelligence and Microsoft AI; and Standard Bank uses Celonis Process Intelligence to transform CIB Operations — from optimizing cross-border payments to enabling transparency, AI, and automation. [2]
What makes this moment different from earlier waves of enterprise software is less the technology itself than the shape of the demand. Organizations are not asking whether AI can do something useful in isolation — they already know it can. The question is whether it can do something useful inside a specific organization, with its specific constraints, at the moment a decision is actually due.
Shah's implementation advice is deliberately narrow: identify a single business decision, isolate its most difficult steps, and test whether AI can improve them. Measuring gains in speed and value then determines whether scaling is justified. [1] The governance question follows the same logic — because agents are assigned specific roles with defined boundaries, organizations can articulate clearly what each agent can and cannot do. [1]
The workforce implications follow a familiar pattern but carry sharper edges in an enterprise context. Automating repetitive, manual work and coordinating tasks across teams could free people to focus on more complex strategic decisions. [1] Realizing that benefit, however, will require investment in workforce education, so employees can understand AI outputs and their implications instead of just handing decisions over to agents. [1]
Shah sees the longest-range opportunity in enterprise-wide planning — finance, workforce management, marketing, and sales. [1] With the right context layer in place, the calculus of how long a consequential decision should take shifts considerably. [1]
The case being made here is not for automation as an end in itself. It is for a redirection of human attention — away from coordination overhead and toward decisions that actually require judgment. Whether enterprises can make that transition without simply replacing one form of opacity with another remains the open question.
Context layers are an architectural choice before they are a product category. Organizations that treat operational context as a foundational requirement — rather than a feature to be added later — are the ones most likely to see AI deliver on the productivity case that leadership teams have been waiting for proof of. [1][2]
Now, maybe [resource planning] should not take that much time, and people are actually making the right choices quickly.Devavrat Shah