From pilot to production: why 40 % of agentic projects will be cancelled, and how to not be one of them

Most agentic AI projects do not fail because the model is not good enough. They fail because production was never designed. The demo works; the evaluation harness, the decision boundaries, the cost model, and the audit trail do not exist.

The four failure modes

Unclear value. The pilot automates a task nobody measured. Without a baseline there is no business case, and without a business case the project dies at the first budget review.

Unbounded autonomy. The agent can do anything its tools allow. Risk and legal teams, rightly, refuse to let it near real customers.

No evaluation. Quality is judged by watching a few sessions. When behaviour drifts — and it will — nobody notices until a customer does.

Invisible cost. Inference spend grows with volume and nobody can attribute it to a use case, so the CFO cannot compare it with the benefit.

What the systems that ship have in common

1. A measured baseline and a value hypothesis agreed before the first sprint.
2. Decision boundaries enforced in code: allow-listed tools, approval thresholds, reversible actions.
3. An evaluation suite with golden data that runs before and after every release.
4. Cost attribution per agent and per step, with a routing strategy that uses the expensive model only where it earns its cost.
5. An operating model for the first 90 days — who watches, who decides, who fixes.

Start with one workflow

Pick a high-volume, rules-heavy workflow with a clear exception path. Put a governed agent into production in twelve weeks. Then scale the pattern, not the pilot.