Decide where it runs first
Before choosing a model, decide where the system will run and where its data may go. The options range from your own cloud tenancy to government clouds, on-premises hardware, and fully isolated networks. The right answer depends on your data, your contracts, and your regulators, and it should be made with your security team.
Treat model services like any other vendor
Read the terms. Enterprise model services can exclude your data from training, and some offer zero data retention. For protected health information, use only services covered by a business associate agreement. For controlled unclassified information and export-controlled data, keep the models inside your boundary.
Give every agent its own identity
Agents should sign in the way people do: through your identity provider, with least-privilege roles, separate credentials, and regular access reviews. When an agent no longer needs access, remove it the same day.
Record everything
An audit trail is the difference between a system you can defend and one you can only hope is working. Record what each agent saw, what it did, which model it used, and who approved it.
Document it for your auditors
Auditors and examiners will ask how the system works, how it was tested, and how it is monitored. Answer those questions in writing before go-live, in the terms of the frameworks you already follow, such as HIPAA, SOC 2, SR 11-7, or NIST SP 800-171.
Questions to ask any AI vendor
Including us.
- Where will our data be processed and stored?
- Is any of it used to train a model?
- Who on your team can see it?
- How is every action recorded?
- How do we pause the system?
- What documentation will you give our auditors?