3.1.2. Channels and Deployment Planning
💡 First Principle: A channel is where the agent lives, and audiences don't come to the agent — the agent goes to them. Channel planning is matching deployment surfaces to where each audience already works, then letting the channel's constraints (auth support, rich content, licensing) validate or veto the choice.
| Channel | Audience | Notes for planning |
|---|---|---|
| Microsoft Teams / Microsoft 365 Copilot | Employees | Natural Entra ID auth; where internal users already are |
| Demo website | Stakeholders during build | Test/share only — not production |
| Custom website (embed) | Customers, public | Pair with the right auth decision from 3.1.1 |
| Custom apps (Direct Line) | Mobile/desktop apps | Developer integration via API |
| WhatsApp and messaging channels | Consumers | Reach without an app; mind rich-content limits |
| Autonomous triggers (event/schedule) | No human at all | The "channel" is an event source — plan monitoring instead of UX |
Two planning consequences deserve emphasis. First, channel and identity interlock: Teams deployment makes Entra ID authentication nearly frictionless (users are already signed in), while a public website makes it impossible for anonymous visitors — so the channel row you pick constrains the identity column you get. Second, capability varies by channel — adaptive cards and rich formatting render differently across surfaces, so an agent designed around interactive cards needs its channels chosen with that dependency in mind, not discovered at rollout.
⚠️ Exam Trap: "Employees should use the agent where they already collaborate" → Teams/M365 Copilot. "Customers without accounts" → custom website with no auth or OAuth. Distractors cross these wires — a public audience on a Teams channel, or SSO promised to anonymous web visitors.
Reflection Question: One agent must serve store employees (in Teams) and customers (on the website) with different knowledge exposure. What does channel planning push you toward — one agent with conditional logic, or two agents sharing components — and which planning principle decides?
