Over the last year, my work with AI has taken place inside enterprise workflows. The kind shaped by long-standing processes, uneven documentation, conflicting rules, and data patched more often than maintained. People move through this environment, off script, with judgment, habit, and quiet adjustments that don't always land in any official process.
Specifically, the AI models I am referring to support tasks inside those workflows. Leaned on to retrieve information, interpret context, and produce structured output based on whatever the system provides. They shine when the environment is steady. When it is not, the shift in output shows up immediately.
That pattern has been consistent across the things I have worked on this year. These models surface the true shape of the environment. They do not pick up on all those informal adjustments people make every day, or compensate well for gaps or reliably smooth over contradictions.
In a sense…
They function like a mirror, showing the system without the adjustments people make to keep it all actually running.
Once that reflection appears, the results become easier to understand. Scattered sources create scattered answers. Conflicting rules create conflicting logic. Thin knowledge creates thin guidance. It often looks like an AI issue when it is really the underlying structure being shown clearly for the first time.
An MIT study estimated that ninety five percent of enterprise AI pilots are falling short. From what I have seen, the difficulty rarely sits just inside the model. When rules do not agree, and when the logic behind the work is unclear or outdated, any model will reflect that directly. It is not failing. It is reporting.
Speed has been a major focus in the efforts I have worked on, including faster responses, shorter calls, and fewer steps. These gains matter, but they can also hide real gaps. When the foundation is unclear, acceleration simply brings the same problems forward faster. The issues were already there. AI just removes the cushioning people have been otherwise stretching to provide.
Things can really shift when the output is treated as information about the system, not a mistake from the model. When the inconsistencies it surfaces are reviewed directly. When the revealed gaps are used as a starting point for repair. I have watched progress move quickly once the reality of the situation is able to be understood clearly, even in an imperfect state.
These AI models are such a useful diagnostic. They show where things hold and where they fall apart. The view can be uncomfortable, but the clarity is powerful. And when that clarity is treated as the first step instead of an inconvenience, meaningful progress tends to follow.
AI exposes. Interesting work begins after that.
