Ozgur Karahan AI Solution Engineer at Microsoft

Enterprise AI architecture

Models matter. What surrounds them decides whether agents get used.

A better model arrives every few weeks, and that matters. But in real enterprise projects, an agent succeeds or fails on everything around it: people talking to it, finding it in Teams or Copilot, reaching the company's workflows and data, acting as the person who asked, and staying under rules everyone can see.

Those five topics matter more every month. They are what I build and test, in the open.

Not covered here, yet: cost control and the value of each token. They are coming fast too.

Let's connect Live now: Oz Radio

01 · The new channel

Voice

Voice is becoming the way we talk to agents.

Typing was the first interface. Speaking is faster and more natural, and it is spreading quickly: customers call voice agents, and even developers now talk to their coding agents. A voice agent has to listen, act, and know when to hand over to a person.

  • Voice Live
  • Speech
  • Phone
  • Avatars

Go deeper

Demonstration scope. Horizon Famille is a real recorded phone call with mock CRM, ERP and knowledge-base services; it is not a production deployment.

02 · Where users are

Distribution

Bring agents to where people already work.

An agent nobody finds is an agent nobody uses. Build it once and publish it where users already are, in Microsoft Teams and Microsoft 365 Copilot, with the same identity and governance behind every channel.

  • Teams
  • Microsoft 365 Copilot
  • Foundry publishing

Go deeper

03 · Workflows and data

Context

An agent is only as good as the company knowledge it can reach.

Bring the company's knowledge to the agent in three ways: MCP servers for workflows and data in systems such as Salesforce, ServiceNow or SAP; knowledge bases; and retrieval (RAG) over company documents. Each choice trades freshness, cost and control differently.

  • MCP
  • RAG
  • Knowledge bases
  • Citations
  • Salesforce
  • ServiceNow

Go deeper

  • ServiceNow meta-tool

    An MCP server: three generic tools (discover, query, write) cover ServiceNow's data model at runtime.

  • Salesforce meta-tool

    A few generic tools discover Salesforce objects at runtime instead of one tool per object.

  • Simple Order Agent

    An order workflow through MuleSoft MCP and A2A, with the Claude Agent SDK.

  • Enterprise Document RAG

    Private work: document intelligence and retrieval over company documents.

  • How DocWiki works

    Foundry agents, MCP and Agent Skills on Azure Container Apps.

04 · From the voice to the system

Identity

The agent acts as the person who asked.

Identity is not a layer; it runs through all of them. Agents take more and more actions for users, in more and more systems, so the user's identity has to travel from the voice that asks, through Teams or Copilot, all the way to Salesforce or ServiceNow. Each system then applies the authentication and authorization rules it already has, instead of rebuilding them in middleware or hiding everyone behind one service account.

  • Entra ID
  • On-behalf-of
  • API Management
  • MCP

Go deeper

These samples are not audited production guarantees. Read the evaluation boundaries before adapting a pattern.

05 · Across every layer

Governance

Every agent, tool and skill needs an owner and a rule.

Governance is not a layer either; it applies at every one. As agents multiply, so do their tools and skills: which channels an agent is published to, which workflows and data it may touch, which skills it carries, where it runs, and where a person must decide. Agents that code or browse also need their own isolated compute, a sandbox per session rather than a shared server.

  • Foundry
  • Sandboxes
  • Approvals
  • Observability

Go deeper

Bring a workflow. Let's map its building blocks.

How people will talk to the agent, whose identity it carries, what it must know, who governs it, and where it will live. I'm glad to talk it through.

Connect on LinkedIn Explore GitHub