An AI feature is not production-ready because the model answered the demonstration correctly. Production means the feature behaves acceptably with unfamiliar inputs, changing data, provider failures, malicious content, different languages, and real people who interpret confidence as authority.
The engineering work around the model determines whether the feature is dependable: evaluation, observability, fallbacks, permissions, cost controls, human escalation, and a release process that treats prompts and tools as versioned software.
A model response is an output. A production AI feature is an operating system around that output.
Define success with examples
Replace ‘good answers’ with a reviewed test set. Include normal tasks, difficult edge cases, insufficient evidence, conflicting sources, sensitive requests, and the languages used by customers. Each case needs expected properties: required facts, prohibited claims, acceptable tools, citation rules, latency, and whether escalation is mandatory.
Evaluate the complete system
Model quality is only one layer. Retrieval may choose the wrong document. A tool may reject valid input. A prompt may omit a policy. A workflow may retry an irreversible action. Evaluate end-to-end traces and classify failures by component so the team can fix the cause rather than adjusting the prompt for every symptom.
Make uncertainty visible
The interface should distinguish a grounded fact, a generated suggestion, and an unavailable answer. When evidence is missing, the safest output may be a clarifying question or a handoff. Users need access to supporting sources and a simple way to report a problem. Confidence should be earned through transparency, not simulated through fluent language.
Design provider and tool fallbacks
A fallback is not automatically another model. The system may retry safely, switch provider, return retrieved sources without synthesis, save a draft for later, or route the task to a person. Choose fallback behaviour according to impact. A customer-support suggestion can degrade gracefully; a financial or production action should stop when required controls are unavailable.
Observe quality, cost, and behaviour
Track latency, token and tool cost, retrieval results, model and prompt version, tool calls, approval outcomes, errors, escalation, and user corrections. Protect private information through redaction and retention rules. Dashboards should answer whether the system is completing useful work and where humans are repairing it—not only whether the API is responding.
Control releases
Prompt, model, retrieval, tool, and policy changes should pass evaluation before release. Use small rollouts, compare versions, and retain a rollback path. A model upgrade can improve general reasoning while changing structured output or tool selection. Treat it like a dependency change with behavioural consequences.
Set an autonomy budget
Define which actions may happen automatically, which require approval, and which are never available. Add limits for steps, duration, spend, data access, and external destinations. The system should not discover its own authority while executing the task.
The production gate
Before launch, require an owned evaluation set, tested failure paths, explicit permissions, observable traces, cost limits, privacy review, human escalation, incident ownership, and rollback. If those controls feel excessive, the feature probably needs less autonomy rather than less engineering.