The demo lies by omission
Every agent looks great on the ten examples you tested in front of the client. The problem shows up on example eleven: the document with a different layout, the ambiguous question, the case nobody anticipated. In a pilot, that's an anecdote; in production, it's a wrong order, an angry customer, a wrong number in a report that goes to leadership.
What separates a useful agent from a pretty demo
- Measured evaluation: a set of real cases with an expected answer, run on every change. Without it, "it got better" is just an opinion.
- Decision limits: the agent suggests, a business rule in the backend validates, and an irreversible action goes through a person.
- A trail: you can see why the agent answered that way — which piece of the knowledge base it used, what the prompt was.
Where agents pay off today
Search over the company's knowledge base (contracts, procedures, history), document field extraction, triage and first response in internal support. These are cases with high volume, manageable error tolerance, and a person in the loop.
Where it still doesn't pay off
Anything where the cost of a mistake is high and the mistake is hard to detect. There, the agent works as a copilot — it speeds the person up, it doesn't replace the decision.
More on how we think about applied AI is in What we do.