Giving an AI teammate broad access because the role is “internal” is not a safety model. Internal work still includes customer details, pricing, contracts, credentials and financial records.

A useful starting exercise is an approval matrix. It turns a vague discussion about trust into specific permissions for specific actions.

Four levels of authority

Use four levels for every task in the role:

LevelWhat the teammate can doExample
ReadRetrieve only the approved inputsRead open quotes in one pipeline
DraftPrepare output without changing anythingDraft a routine follow-up
ProposeRecommend a record change or assignmentSuggest a new follow-up date
ActComplete a pre-approved low-risk actionMark a reviewed task complete

Sending external email, changing commercial terms, granting access and moving money should not become “Act” merely because earlier drafts looked good.

Match permissions to the role

If the job is quote follow-up, the teammate does not automatically need your entire mailbox, every Teams channel or every ERPNext module. Connect the specific shared inbox, channel and record set needed for that queue.

The same rule applies to files. A role that collects onboarding documents needs a defined destination and checklist, not unrestricted access to the company drive.

Make approval informative

An approval button is weak if the reviewer cannot see what they are approving. Include:

  • The record and current values.
  • The proposed message or change.
  • The source used to prepare it.
  • The reason the item entered the queue.
  • Any uncertainty or conflicting information.

Avoid batches that encourage rubber-stamping. Review should be quickest on genuinely routine work, not on items with missing context.

Keep an activity trail

Record the assignment, source records, proposed action, approval decision and completed change. This gives the team a way to answer “what happened?” without relying on a generated summary alone.

Also decide how to revoke access and stop scheduled work. An off switch is part of the design, not an emergency afterthought.

Test with uncomfortable examples

Before launch, try duplicate contacts, ambiguous requests, prompt injection inside an email, outdated templates and a user asking for access outside the role. The correct result is refusal or escalation where appropriate.

MSPE’s managed AI teammate service starts with scoped access and supervised output. Your exact guardrails should reflect your real systems, data and obligations, so review them with the people who own those risks.