The prototype hides the distribution
A demonstration usually contains cooperative inputs selected by its maker. Real work contains missing context, unusual formats, conflicting instructions, malicious content and cases nobody remembered to mention. Build an evaluation set from representative and difficult examples before tuning prompts around memorable failures.
Define useful quality dimensions separately: factual correctness, completeness, format compliance, appropriate refusal, tone and latency. A single pass rate can conceal a dangerous weakness.
Reliability is a system property
Models sit inside products. Retrieval quality, permissions, tool behaviour, timeouts, retries, version changes and user expectations all affect the result. Store enough structured evidence to investigate failures while minimising personal data and respecting retention requirements.
Design explicit fallbacks. The product should explain when it cannot complete a task, preserve the user’s work and provide a safe route forward. Silent invention is not a graceful degradation strategy.
Operating cost includes attention
Token cost is only one line. Add evaluation, observability, incident handling, vendor review, human approval and the cost of correcting output. Measure cost per successfully completed outcome, not cost per model request.
Latency also changes behaviour. A slower, more capable model may reduce completion because users abandon it. A faster model with targeted tools and validation can create the better product.
Production begins with ownership
Name the person accountable for quality thresholds, data decisions, provider changes and incidents. Define what triggers a rollback and how people perform the task when the AI path is unavailable.
A prototype should therefore test the operating model as well as the output. The honest conclusion may be to narrow the use case, keep permanent review, use conventional automation, buy an existing tool or stop. That is successful discovery.