Making the case for context-aware AI

Devavrat Shah, chief scientist of enterprise AI at Celonis and a longtime MIT professor of AI, makes the case in an MIT Technology Review Business Lab episode that enterprise AI's next step depends less on model capability than on operational context. The episode, hosted by Megan Tatum and produced in partnership with Celonis, frames the industry as being at a crossroads after years of pilots: leadership teams want proof that scaling AI affects the bottom line. Shah's answer is what he calls decision intelligence, in which agents take on individual tasks inside a larger decision process while understanding the role each task plays and the business conditions around it. His shorthand for the alternative is that this cannot be done in a vacuum.
Shah splits enterprise decisions into two rough kinds: strategic decisions that are episodic, and operational decisions that must be made at high frequency. Both run through processes with ordered steps, and people coordinate those steps and occupy roles within them. He argues agents can supply capability at that atomic level, augmenting or supporting parts of a decision and sometimes automating painful steps. Knowing which task or role to assign requires context, which is why he couples context with AI under the decision-intelligence label.
Asked about commercial risk when context is absent, Shah offers two failure modes. The first is cost: a sufficiently capable model might eventually work out a narrow decision by asking many questions the way a person would, but reaching that point burns enormous compute on a simple task. The second is error: the model misses an important constraint or guardrail and makes a mistake. Some mistakes are cheap, like a restaurant recommendation the diner does not enjoy; others are very expensive, such as a financial forecast that leads to the company's doom.
Celonis's route to supplying that context comes out of its history in process intelligence, helping organizations understand how their processes actually run, where they break down and how to re-engineer them. Shah describes a context layer, which the company calls a context model, that sits between two other layers. Below are systems of record, workflow tools and the event logs those tools generate. Above are business experts, either reading information directly or receiving output from agents operating on the context layer. The context model is meant to do three things at once: abstract the layer below so the right pieces of a process can be agentified; bring what Shah calls roled agents to those pieces in a structured, guardrailed way; and be agent-first, so other agentic infrastructure already in place can talk to it naturally. Outputs are meant to reach humans or experts for review and approval, and sometimes automation.
On readiness, Shah is notably relaxed about the state of enterprise data. The systems of record and workflow tools already exist, he says, and need not be complete or perfectly clean, because the context model's architecture is built to connect to them and make sense of what is there. The layer above is under Celonis's own control, so it is not especially complicated. The critical gate, in his account, is classical information security review: opening up a system of record, workflow tool or event logs to a vendor requires establishing that the vendor's software can be trusted inside the organization.
Shah's implementation advice is deliberately narrow. Pick one business decision, map the process around it, break out each step, take the hardest and most crucial step and try agents there. Then measure the speed and value gained to judge whether the application is worth scaling. Governance follows from the roled structure: because agents are prescribed for a specific task within a role, an organization can state clearly what that role may and may not do, and Celonis prescribes certified role agents for this purpose, which Shah likens to having a clear contract. The third pillar is people. Where experts remain in the loop, they need to understand AI outputs, how to work with them and what their implications are, which calls for both AI education and tool training.
On the workforce, Shah expects a shift rather than a simple reduction. Repetitive, manual, grungy work and coordination-heavy tasks are the obvious candidates for automation and faster throughput, which he says reduces lost opportunity cost and improves planning. Strategic decisions, the kind organizations often lack bandwidth to think through, are where he expects more human and expert engagement, with AI helping compare options while the organization retains the choice. He also expects new kinds of work to appear that are hard to predict in advance, and stresses that re-education must happen at the pace of the change itself.
Looking ahead, Shah groups the opportunity into three categories: repetitive manual labor-intensive work; coordination work, the email-and-spreadsheet kind that an information workforce spends much of its time on; and strategic support that augments human decision-makers. Enterprise-wide planning across finance, workforce, marketing, sales and product is where he sees the largest shift, arguing that annual planning cycles need not take as long and that people can make the right choices quickly.
Why it matters: Teams building agents for business processes should read this as a nudge toward investing in process and context plumbing, guardrails and evaluation before scaling autonomy, since the cost of missing constraints shows up as both compute spend and bad decisions. The layered design Shah describes, with a context model between record systems and agentic tooling, is a template developers may increasingly be asked to integrate with, and it makes information security review a gating step rather than an afterthought. The claim that messy data is tolerable is worth testing rather than assuming, since it is the vendor's characterization of its own product; in practice, the quality of process abstraction likely bounds how much any agent can safely automate.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Enterprise AI can generate recommendations, automate tasks, and help organizations make decisions faster. But without an understanding of how a business actually operates, even powerful models can miss the constraints and relationships that determine whether a decision is the right one. As companies look to move AI from experimentation into everyday operations, the challenge ahead…