Field guide
Four questions before an agent gets access
Memory, permissions, review and portability determine how much control a person keeps.
1. What information can it access?
Make the data boundary explicit: approved documents, a specific workspace or a narrow set of records. A system that can retrieve a document also needs a rule for whether a particular user should see it. Source content can contain misleading instructions; retrieval should supply evidence rather than grant authority.
2. What can it do?
Reading, drafting, sending and changing a record are different permissions. Start with the smallest action set that serves the task. A useful product makes the proposed action visible before an important change is committed.
3. How can a person inspect and correct it?
Look for source links, a record of meaningful changes and a path to correct a mistake. A percentage score without a defined evaluation set does little to explain the failures a real user will encounter.
NIST’s AI risk framework ↗ provides a broader reference for evaluating the use context and managing failures.
4. What happens if the provider changes?
Ask how to export documents, conversation records that belong to the user, and the workflow’s configuration. Separate the lasting business knowledge from the current model provider. Costs, retention policies and migration effort belong in the design discussion.