Enterprise AI Needs Business Context to Avoid Costly Mistakes
Celonis chief scientist argues that AI agents require operational understanding of workflows and constraints to make useful decisions in complex business environments.
Last verified:
The Context Problem in Enterprise AI
Generalist large language models excel at pattern recognition across broad domains, but enterprise AI agents operating within specific business processes face a fundamentally different challenge: they must understand not just what to do, but whether a proposed action aligns with organizational constraints, compliance rules, and workflow dependencies. According to MIT Technology Review AI, Devavrat Shah, chief scientist of enterprise AI at Celonis, argues that this operational awareness is the missing link between AI pilots and reliable production systems.
The risk is concrete. An AI agent optimizing for speed might recommend a supply-chain decision that violates a contractual obligation. Another might automate a task without recognizing that a downstream approval process depends on human judgment at a critical juncture. Shah emphasizes that “you can’t just do that in a vacuum”—isolated optimization often produces decisions that appear sound in isolation but create coordination failures or resource waste when viewed through the lens of actual business operations.
Building Context Into AI Workflows
Shah’s prescription is neither to reject AI automation nor to deploy agents blindly, but to architect systems that surface organizational context before decisions are made. According to MIT Technology Review, the solution involves a “contextual layer” that helps agents understand workflows, operate within defined guardrails, and produce recommendations for human review rather than autonomous execution.
The practical approach begins with specificity: organizations should identify a particular business decision—not AI adoption in general, but a concrete task like workforce scheduling or purchase-order routing—and diagnose which steps would benefit most from acceleration or consistency. This incremental approach allows teams to measure whether the AI intervention delivers measurable gains in speed or cost before committing to broader rollout.
Workforce Implications and Required Investment
Shah sees potential for substantial productivity gains if enterprise AI can automate repetitive manual work and coordinate tasks across functional silos, freeing employees to focus on higher-level strategic decisions. However, realizing this benefit requires investment in employee education, so workers can interpret AI outputs and their implications rather than delegating decisions wholesale to agents without understanding the reasoning.
The timeline extends beyond routine operations. Shah forecasts opportunities in enterprise-wide planning domains—finance, workforce management, marketing, and sales—where AI could help organizations evaluate alternatives and make consequential decisions faster, potentially reducing planning cycles that currently consume disproportionate time.
Why This Matters
The distinction Shah draws between generic AI capability and contextually grounded AI agents addresses a widening gap between CEO expectations and pilot results. Organizations that treat enterprise AI as a plug-and-play capability often discover that models trained on general corpora make recommendations that violate domain-specific rules or miss interdependencies that domain experts take for granted. By contrast, teams that invest in mapping workflows, defining constraints, and building feedback loops for continuous refinement are more likely to achieve the measurable ROI that will accelerate adoption beyond the pilot phase. This framing suggests that the next wave of enterprise AI success will be determined not by model size or latency, but by architecture that embeds organizational knowledge into agent decision-making.
Frequently Asked Questions
What is decision intelligence in the context of enterprise AI?
According to Devavrat Shah, decision intelligence is the ability of AI agents to understand their role within larger decision-making processes and the business conditions surrounding them, rather than operating in isolation.
Why do AI agents need operational context?
Without understanding business workflows and constraints, AI agents risk overlooking important guardrails, making costly mistakes, or consuming excessive computing resources as they attempt to determine what information they need.
How should organizations approach deploying enterprise AI?
Shah recommends starting with a specific business decision, identifying its most difficult steps, testing whether AI can improve them, and measuring gains in speed and value before scaling.