Two years into the agent era, the pattern is clear: the enterprises getting returns aren't the ones with the most pilots — they're the ones that treated agents as systems engineering rather than model demos.
Where agents are winning
The consistent winners share three traits: high volume, document-heavy inputs, and decisions that are rule-informed but exception-rich. Loan origination checks. Invoice-to-PO matching. Service-ticket triage. In each, the agent doesn't replace judgment — it clears the 80% that never needed judgment, and assembles context for the 20% that does.
One of our lending clients cut approval turnaround 68% not by letting a model approve loans, but by letting agents do everything except approve loans: extraction, verification, cross-checks, and routing, with every decision logged for the risk committee.
Where agents fail
Agents fail where their sponsors skipped the boring questions. No evaluation set, so nobody notices quality drift. No cost ceiling, so the token bill scales with enthusiasm. No action allowlist, so security kills the project at review. The failures are rarely model failures — they're missing engineering.
The architecture that survives
Production agent systems converge on the same shape: a planner bounded by explicit tool contracts; confidence-based routing between straight-through and human lanes; an event log that makes every run replayable; and an evaluation harness that gates releases the way tests gate deploys. Build that skeleton first and the model choice becomes almost boring — which is exactly the point.
Where to start
Pick one queue your operations team already resents. Measure its baseline honestly. Ship an agent with a human lane and an audit log. When the metric moves, you'll have something better than a strategy deck: evidence.