POC to prod, signed
The gap between a working AI prototype and a production deployment is wider than most teams expect, and most of the risk your board worries about lives in that gap. Latency that doubles under real traffic. Costs that scale faster than usage. Drift that nobody notices for weeks. Audit gaps your regulator finds before you do. None of this is exotic - it's the basic operational hygiene that any other production system would have.
The shift is from "we got it working" to "we can run it." Latency budgets in CI. Eval harnesses tied to PRs. Observability that shows per-request traces, cost-per-call, and the model's actual decisions in the same tooling your platform team already uses. Redaction so PII never reaches a model provider. Fallback paths so your business stays up when a model is down. Signed audit on every state change.
This is the part of the work that doesn't demo well. It's also the part that determines whether the program is still running in eighteen months. We deploy into your VPC, stay paged for ninety days, and exit when your operators are the ones solving incidents - using a runbook they helped write.

Three ways this shows up in production.
Latency budget
P50, P95, P99 baked into the eval harness. CI fails when it gets slow.
Eval harness
Golden sets, regression gates, drift tracking. Tied to PRs.
Observability & on-call
Per-request traces, cost-per-call, runbooks, paged 90 days.