Where AI actually belongs in an operational workflow
The useful applications are narrow, bounded and reviewable. The disappointing ones are general.
5 min read
General assistants disappoint in operations
A general assistant bolted onto a business system tends to produce a specific pattern: an enthusiastic first week, then declining use. Not because the model is weak, but because the task was never defined well enough for anyone to tell whether the answer was good.
Operational work has correct answers. That is what makes it operational. A tool that produces plausible answers to questions with correct answers creates verification work rather than removing it.
Bound the task, then judge it
The applications that hold up are narrow: extract these fields from this document type, classify this request into one of these categories, draft this recurring reply for someone to approve. Each has an observable success condition, so quality can be measured rather than felt.
A bounded task also fails safely. When a classifier is unsure, it can escalate. When a general assistant is unsure, it answers anyway.
Keep a human where the consequence is
The question is not whether a human reviews the output. It is whether a human reviews the output at the point where something irreversible happens.
Drafting a message can be automatic; sending it to a customer may not be. Proposing a credit can be automatic; issuing it is a financial event. Putting the approval step exactly at the consequence — rather than everywhere, or nowhere — is what makes automation acceptable to the people accountable for it.
Earn the next step
Start with the task that is high-volume, low-variance and currently done by hand. It is rarely the most exciting candidate, and it is usually the one that pays for itself first.
Broader autonomy is something to earn from a measured track record, not something to configure on day one.