We've been dealing with this during our own rollout, and the biggest realization is that a prompt is not a security boundary. You can tell an LLM "don't look at private data," but that's just a suggestion. Real governance happens at the identity layer.
Why this actually matters for the workflow #
When we move from a basic chatbot to a real LLM agent, the agent starts chaining tools. If an employee asks for a meeting brief, the agent might read an email, pull a product doc, and then update a lead status. In a naive setup, the agent uses one master API key for all three. That's a disaster waiting to happen because those three actions have totally different risk profiles.
If you don't have a proper identity map, your audit logs become useless. You'll see that "The Agent" performed an action, but you won't know if it was acting on behalf of the user or using its own system-level authority.
Breaking down the Foundry identity confusion #
There are four concepts in Microsoft Foundry that everyone on my team kept mixing up. If you're trying to build a practical tutorial for your devs, make sure you clear these up first:
End User Identity: This is the person actually logged in. It's the "who" initiating the request.Agent Identity: This treats the agent as its own principal. It allows the system to identify and audit the agent independently from the user.Managed Identity: This is the technical plumbing that lets the agent authenticate to Azure resources without us having to hardcode secrets or rotate keys manually.Blueprints: These act as the template for the agent's identity and permissions, ensuring that when we deploy a new version of an agent, it doesn't suddenly have different access rights than the last one.
The goal is to move toward a "least privilege" model. Instead of one giant key, the agent should switch contexts based on the tool it's calling.
For those of us doing the actual deployment, the struggle is usually convincing the business side that we need to spend time on this now rather than "fixing it later." But seeing a 200 OK response in the logs for a data leak is way scarier than a 403 error during a sprint.
If you're building this from scratch, I'd suggest mapping every single tool the agent uses to a specific identity requirement before you even write the first line of the agent's system prompt.
[Next Selling niche API documentation is a low-overhead way to hit →](/en/threads/5713/)
[these real-world AI monetization case studies](https://tanyan888.com/), with plenty of directly applicable cases.