Put safety classification and approval gates before the LLM

Prompt instructions cannot carry the full safety burden of a vertical AI product. A model can misunderstand a request, follow hostile content, expose sensitive context, or prepare an external action the user did not intend. Product controls need to surround the model call.

Classify risk before generation

Digital Wellness Academy uses MAIA to classify distress before a coaching model sees the message. The vertical declares the classifier in its manifest. Commercial products can select a different path because their risk model differs.

The same architectural pattern applies outside health. A sales product can classify an action as internal drafting or external communication. A finance workflow can separate analysis from a transaction. A support product can route security and account-access requests to a controlled process.

Minimize sensitive context

Redact or exclude data the task does not require. Retrieval should apply tenant, role, purpose, and product filters before returning context. Logs should record operational evidence without copying sensitive text when a hash, identifier, length, or policy result will serve the audit need.

Place approval before side effects

A draft can remain inside the product. Sending email, updating a CRM, changing access, publishing content, or making a purchase creates an external side effect. The product should surface the exact action and destination before execution when policy or user expectations require approval.

Design conservative failure behavior

A classifier outage, retrieval failure, or provider timeout needs a defined product response. The system can pause the action, show a safe fallback, route to a human, or restrict the response. Silent degradation creates the largest gap between product claims and product behavior.

Test controls without paying for model calls

State machines, policy functions, permission checks, and failure routing can run in deterministic tests. Model evaluations then focus on the behavior that requires inference. This split improves coverage and keeps the safety contract visible.