Ask how they separate AI from application logic
A robust architecture keeps business rules, authorization and irreversible actions outside the model. Ask where deterministic logic lives and which decisions are delegated to AI.
If the answer is “the prompt handles it,” risk is high. Prompts guide behavior; they do not replace access control, transactions or policy enforcement.
Ask for an evaluation strategy before launch
The team should describe how they will measure quality on representative tasks. That includes expected answers, tool-use accuracy, groundedness, refusal behavior and known failure cases.
Without evals, every model or prompt change becomes subjective. A professional delivery process turns failures into regression tests.
Ask how your data is protected
Clarify what data is sent to model providers, how tenant access is enforced, what is logged, where embeddings are stored and how secrets are handled.
The answer should map to your threat model and data classification rather than vague statements about “enterprise security.”
Ask what happens when the model fails
Providers can be unavailable, slow or rate-limited. Ask about timeouts, retries, provider abstraction, fallback models and deterministic fallback paths.
A critical workflow should not become unusable because one API returns 429 or takes 30 seconds.
Ask how production behavior will be observed
You need traces for model version, prompt version, retrieval, tool calls, latency, validation failures and final outcomes.
If the team cannot explain how a bad answer will be debugged in production, the system will be expensive to operate.
Ask who owns the architecture and release process
Clarify code ownership, infrastructure ownership, CI/CD, documentation, secrets, environments and handover.
A serious partner should leave you with an operable product, not a dependency on one developer’s laptop or hidden account.