AI & agents
What we learn moving organisations onto the Microsoft cloud. Method, not opinion.
One agent is an integration. A hundred is a governance problem.
Agents got production identities before they got a stable way to record what they did. Microsoft Entra Agent ID is generally available, with sponsors and expiry dates. The OpenTelemetry conventions for agent traces are still experimental. That gap is not a detail: it decides what you can responsibly put into production this year, and it turns an AI question into a governance one.
Agents decide. Code executes. Knowing the line is the job.
The promise is that agents replace applications. In the systems we build, they do not. An agent is very good at understanding a request, weighing options and choosing a tool. It is the wrong place for a payment, a tax rule or a permission check. This is where the line falls, and why putting it in the wrong place is expensive.
AI solved syntax. It did not solve judgement.
Three independent measurements, taken between 2025 and 2026, all point the same way: AI-generated code passes over 95% on syntax and 55% on security, and that second number has not moved in two years. The part you write got cheap. The part you decide did not.
Can your AI assistant read documents the user cannot see?
A badly built RAG searches with the application's identity, not the asker's, and the assistant answers with what the person should not see. The fix is not a filter someone can forget to write: it is passing the user's token and letting Azure AI Search enforce permissions inside the engine.
What to put in order before switching Copilot on
Copilot does not create permission problems. It makes the ones already there findable, and you no longer need to know what you are looking for.
Nothing matches your search.