Pick use cases by cost of being wrong
A good first production use case has three properties: the task is repetitive enough to matter, the output is checkable, and a mistake is recoverable. Summarising service history, drafting responses, classifying documents and retrieving policy answers all qualify. Autonomous pricing or credit decisions do not, at least not first.
Framing the decision this way also settles the governance conversation early. Risk, legal and compliance teams block projects that arrive as finished surprises; they tend to cooperate when the containment strategy is part of the original design.
Data access is the real project
Most AI work turns out to be data work. The model needs current, permissioned, well-described data, which means resolving identity across systems, agreeing what a customer record is, and enforcing the same access rules the source systems enforce.
Retrieval quality follows directly from this. Poor answers are usually a retrieval problem — stale content, missing metadata, chunking that destroys context — rather than a reasoning problem, and swapping models will not fix it.
- Resolve identity and entity definitions before building retrieval
- Carry source-system permissions through to the AI layer, never around it
- Keep a freshness contract for every data source the system depends on
Evaluation is what makes it shippable
A pilot is judged by demo. A production system is judged by a test set. Build a graded set of real cases with expected outcomes, run it on every prompt, model or retrieval change, and track regressions the way you track failing unit tests.
Pair that offline evaluation with online signals: escalation rate, human edit distance on drafted output, task completion, and the specific failures users report. Without both, a change that feels better in review can quietly degrade the system.
Design the escalation path and the run cost
Every production AI system needs a defined boundary — what it may do alone, what needs confirmation, what it must hand to a person, and how that handover preserves context. Systems without a clean escalation route erode trust the first time they fail in front of a customer.
Run economics deserve the same attention. Token spend, latency budgets, caching, model routing by task difficulty and monitoring belong in the design, because an AI feature that works but costs more than the labour it replaces will not survive its first budget review.