Can an embedded AI engineer build across departments without getting access to payroll, HR files, and all our customer data?
Yes. Design access around the workflow being built, not around the engineer's ability to request broad administrator permissions. Being embedded should mean understanding your operations, not having unrestricted visibility into the business.
Start by mapping what the application needs to read, what it can change, and who should see each result. A service scheduling tool might need employee availability without needing compensation or personnel records.
Use these boundaries:
- Separate development and production environments, using synthetic or appropriately de-identified data where practical.
- Give the engineer named, limited-access accounts rather than shared administrator credentials.
- Keep integration secrets out of prompts, source code, and browser-visible application code.
- Enforce permissions on the server or database, not just by hiding screens.
- Document what data reaches an AI provider, including retention and contractual restrictions.
- Remove temporary access when the work ends.
When building with Lovable or Replit, verify the actual configuration of the application and its connected services. The choice of development tool does not establish that the finished system has appropriate access controls.
Use EOS as the backbone for accountability: identify who approves access and who owns the affected process. Put unresolved access decisions on the Issues List rather than letting the engineer make business-policy decisions by default.
My recommendation is to approve the smallest useful access scope first. Expand it only when a specific requirement justifies the change. Sensitive workflows may also require security, legal, or compliance review before live data is introduced.
Category: Forward-Deployed Engineering