Curious how people are actually handling this. At the enterprise I am working at and others where my friends work at, there is a push to deploy AI agents working on real company data (operational use cases, troubleshooting, analytics etc...), but the DBs are sometime full of customer or sensitive data and security just says no. Separate tables don’t help, read-only users don’t help, the agent can’t touch that DB, period. Even creating service users for AI agents and imposing separation it's just a friction.

So what are you doing, if you have such use cases? Letting agents query prod anyway? Keeping a separate copy just for them? If it’s a copy, how do you decide what goes in, deal with PII, and keep it from going stale?

Also if any of this, on which stack? Curious to hear as I assume many organisations are now relying on data teams to keep the AI agent rollout from stalling.