Connecting a model to tools changes the consequence of an incorrect answer. Design the permission and review model before giving the system authority to act.
1. Separate assistance from action
Begin with a defined task and read-only information access where practical. Distinguish drafting a recommendation from changing an operational record. Document who can authorize actions and which decisions always remain with a person.
2. Enforce permissions outside the model
Apply user and service identity at the retrieval and tool boundaries. Use narrowly scoped operations and validate inputs. A text instruction in a prompt is not a substitute for access control. Retrieved documents and tool output should be treated as untrusted data.
3. Evaluate more than answer quality
Include unsupported claims, access violations, misleading source content, tool failures, and escalation behavior in testing. Preserve the evaluation set and results with the release. Define acceptable behavior for uncertainty, including declining to answer or asking for review.
4. Plan for changes after release
Model versions, source corpora, and tool interfaces evolve. Monitor quality and operating cost, record consequential actions, and maintain a rollback path. Expand authority only after reviewing representative evidence and updating the organization’s approval and incident processes.
helixoid