Start with the job
The useful question is not where can we add a model. It is which part of a user's workflow is expensive, repetitive, or hard to navigate, and whether a model can help without weakening correctness.
That usually leads to a narrow capability: classify, extract, suggest, summarize, or choose from a known set of actions.
Keep the model inside a boundary
The model can interpret intent, but application code should own permissions, validation, calculations, and side effects. Structured outputs and typed tools make that boundary testable.
A fallback is part of the feature. If confidence is low or a dependency is unavailable, the system should return to a manual path rather than improvising.
Measure usefulness
Completion rate, correction rate, and time saved are better signals than model cleverness. A smaller model with predictable behavior often creates the better product.
A bounded AI workflow
The model should sit inside a workflow with explicit stages. The application owns authentication and permissions before the model sees the request. It validates the model's output before a tool can mutate state, and it records the final outcome for evaluation.
flowchart LR
A[User intent] --> B[Auth and context]
B --> C[Model interpretation]
C --> D[Schema validation]
D --> E{Safe action?}
E -->|No| F[Ask for clarification or human review]
E -->|Yes| G[Domain command]
G --> H[Durable result]
H --> I[Feedback and evaluation]
Validate before acting
A structured response is not automatically a correct response. Validate enum values, resource ownership, numeric ranges, and whether the requested operation is allowed in the current state. The command handler should accept the same typed input whether the request came from an AI model or a normal button.
const action = ActionSchema.parse(modelResponse);
await permissions.assertCanRun(user, action);
const result = await domainCommands.execute(action);
await auditLog.record({ userId: user.id, action, result });
Design for uncertainty
Low confidence should change the workflow, not just the wording of a warning. Ask a clarifying question when the ambiguity is cheap to resolve. Route to a person when the cost of a wrong action is high. Fall back to a normal form when the model or provider is unavailable.
The strongest AI feature is often the one that knows when not to act.