Manish Garg is the cofounder and chief product officer at Skan.ai, an AI process intelligence platform.
I am sure this scenario seems familiar to most of you who work in enterprise AI: The pilot went well. The demo looked impressive. Then the agent went into production, hit its first real exception and did something confidently wrong.
Don’t blame the model. Trying a newer one doesn’t work, either. Adding guardrails is good, but still not enough. Writing a better prompt doesn’t really fix it. None of it closes the gap, because the gap is not in the model. It is what the model was given to work with.
We have spent the past several years watching enterprise processes and workflows, not by reading about them, but by observing them directly through computer vision and machine learning across billions of executions in Fortune 500 finance, insurance and healthcare.
What that discipline has taught is specific: The way work is documented, and the way work actually runs, are two different things. Agents trained on the first will often struggle with the second.
The Layer Nobody Is Filling Fully
The enterprise AI community has correctly identified that agents need context. Context graphs, enterprise context layers, retrieval-augmented architectures: the direction is right. What I see consistently underestimated is where the context is supposed to come from.
Most agentic AI implementations get data from three sources: documentation, system logs and CRM or ERP records. These capture what the organization says it does, what systems recorded after the fact and what transactions were logged.
Each has value. None captures the reasoning layer that lives between system events: the five-second cross-reference an experienced processor runs before approving an unusual invoice, the informal escalation that happens before anything gets logged or the compressed verification sequence the Frankfurt compliance team uses in the last 10 days of each quarter that the Chicago team never uses at all.
That reasoning layer is where judgment lives, and where most high-value work happens. It is invisible to the sources most organizations use to train their agents.
How A 1% Gap Becomes A 40% Failure
Your data teams learned this lesson the hard way a decade ago: You cannot fix a source data problem downstream. Sophisticated synthesis does not recover decision traces that were never captured.
Compounding of a small matter into a giant mess is the part most teams miss. A modest gap at the observation stage, say 1% of behavioral variation unseen, becomes 5% distortion when patterns are synthesized, 15% percent when structured into training data and 40% agent failure by the time execution hits production. Of course, the numbers are illustrative, but you get the point. The pattern is consistent across enterprises that built automation on fragmentary observational foundations.
An invoice processing team I observed had five distinct execution pathways. Documentation described one. An agent trained on that documentation handled routine cases adequately. In week 13 of the quarter, when override pathway volume spiked above 20% of invoices, the agent flagged urgent approvals for human review and applied standard timelines to time-sensitive exceptions, producing exactly the bottleneck the automation was meant to eliminate. The agent was not broken. The source data was.
Seven Dimensions, Not One
As a part of our work over the years, we learned that operational behavior is governed by at least seven context dimensions simultaneously: business rules, role and expertise, time and fiscal cycle, geography, regulatory jurisdiction, organizational dynamics and situational pressure.
Each dimension individually looks manageable. The interaction is where the variation that breaks agents’ lives occurs.
A rough estimate suggests that approvals take two days. Granular observation shows that approvals take four hours for established vendors during mid-quarter with experienced processors in the U.S., and three days for new vendors during quarter-end with junior processors in the EU.
The rough estimation is not wrong, but it is operationally useless. An agent that knows only the average cannot move through the distribution. It fails at exactly the edges where judgment is most valuable.
What Actually Works
The framework for building operational context has three parts, and the sequence matters.
First, we think observation is essential before synthesis. You cannot reconstruct the reasoning layer from logs or records after the fact. It has to be captured directly, as work executes, at the granularity where expert-level judgment becomes visible.
Second, we think the schema should preserve dimensional fidelity. Raw observation produces records; agents need structure. Logically, a semantic model must connect observed context to decisions, decisions to policies and outcomes to feedback. Without such structure, observational richness collapses back into averages by the time it reaches training.
Third, the operational context layer must include an execution-feedback loop that strengthens rather than corrupts. Every agent run produces new behavioral data. If you don’t address it, the system encodes agent errors into the next generation. This design decision determines whether your AI investment improves with scale or degrades as you scale.
I am not going to tell you that this problem is solved. It’s a work in progress. What I will tell you is that this investment compounds. Observational intelligence cannot be bought as a finished artifact or reconstructed from logs after the fact. Enterprises that start now will own autonomous execution capabilities that their competitors cannot replicate within a budget cycle.
The strategic implication is simple. Model capability is converging rapidly across vendors. Orchestration frameworks are commoditizing. The durable differentiator for enterprise AI will not be which foundation model you choose or which agent platform you deploy. It will be the quality of the operational intelligence you trained them on.
That is the layer worth investing in. It is also the one most enterprise roadmaps still treat as somebody else’s problem. Mine would not.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
