We’ve now sat with enough of these engagements, both the ones we started and the ones we were called in to rescue, that the failure modes have settled into a small number of patterns.
The prototype was solving the wrong slice of the problem.
A demo agent that drafts a beautiful first response to a customer email is addressing maybe 15% of the cycle time on that case. The other 85% is the research, the systems lookups, the policy interpretation, the routing, the handoffs, the quality assurance. Agents that move metrics own a wider slice of the work, not a deeper version of one corner.
There was no evaluation harness.
The prototype was graded on the team’s gut feel during the demo. Three months in, every prompt change is an act of faith. The first time the model upgrades (and it will), the team loses confidence in their own system and the project freezes. An eval harness is to an agent what unit tests are to a service. You don’t ship without one. You don’t iterate without one.
The surrounding process never changed.
Operations workflows are shaped around the constraints of the work as it was, not as it could be. An agent that’s faster than a human at one step but then has to wait two days for a manual handoff hasn’t moved the customer-visible metric. The agent’s value is realised by re-engineering the process, not by optimising one step in it.
Governance was bolted on at the end.
Risk and compliance found out about the project two weeks before launch, asked the obvious questions nobody had ready answers to, and the launch got postponed by a quarter. Build the audit log, the eval set, the human-in-the-loop checkpoints and the model-fallback story in the first sprint. The governance team isn’t your opponent, they’re your launch dependency.
No one owned the operational outcome.
The project plan tracked features shipped. Nobody had committed to a number: cycle time, error rate, cost per case, NPS. Without a number, every interesting thing the agent does is “progress”, and the project never has to confront whether it’s working.
The teams that get out of pilot purgatory share a few habits. They start with a process redesign and an outcome metric. They build an eval set before they build a prompt. They get risk and compliance on the team in week one. They expect to throw away a meaningful slice of their first build. And they have someone senior whose only job is to refuse a launch that doesn’t move the number.