Microsoft Foundry Voice Agents: The Phone Layer Is Now a Platform Feature

Microsoft now sells the phone layer of an AI receptionist as a platform feature. Voice agents in Microsoft Foundry Agent Service are in public preview, and the telephony documentation, dated September 25, 2026, covers two ways to put one on a phone number: Teams Phone extensibility, where your Teams tenant keeps the number, and Twilio, where Foundry discovers your voice-capable Twilio numbers and binds an agent to each. Inbound calls come in; outbound calls go out through a call-job API.

Microsoft's docs are careful to say this is preview, "provided without a service-level agreement," and not recommended for production. Take that at face value. But the shape of what's shipping tells you where the market is going, and it confirms something we said in our voice AI receptionist piece: the telephony and speech layers are becoming commodity. The business logic isn't.

What Foundry gives you

Reading the docs, the preview handles the parts that used to take a week:

• Number binding. For Twilio, you connect with an Account SID and auth token, pick the voice-capable numbers from your account, and Foundry creates a binding per number. The docs warn that this replaces the number's existing inbound voice routing, so use a number you intend to dedicate.
• Teams Phone path. Your organization keeps a service number in Teams, routes it through a resource account, and Azure Communication Services carries the media. A Teams or Microsoft 365 admin does most of the setup.
• Suspend without deleting. A binding can be set to suspended, which stops new inbound calls while keeping the configuration. Useful for after-hours or a bad day.
• Real-time models. Coverage of the September announcement lists GPT Realtime, Microsoft's MAI models and custom models as options, with turn-taking and interruption handling built in.

If you already run Teams Phone, the appeal is obvious: the AI answers on the number you already own, and the escalation path to a human is a Teams transfer.

What it doesn't give you

Everything the caller actually wants.

The receptionist we described in AI receptionist ROI books into a real calendar, knows the service catalog, writes to the CRM and escalates with context. None of that is a telephony feature. In that post we also said the thing that pushed booking accuracy from around 70% to above 95% was tailoring the booking logic to the client's actual services, not the platform. That's still the whole game.

Concretely, on top of a Foundry voice agent you still need:

1. A service catalog the agent can reason over. What you offer, how long each takes, who can do it, what it costs, what questions qualify a caller.
2. Calendar integration with rules. Not "find a slot" but "find a slot with a technician who does this service, in this territory, with travel time." Microsoft Graph gets you the calendar; the rules are yours.
3. CRM write-back with matching. Caller's number matches an existing contact or it doesn't; the agent should know which before it asks for a name.
4. An escalation design. Which situations go to a human, how fast, with what summary. The single most-used safety feature in our deployments is the "send to the owner's phone" escape hatch.
5. Logging you can review. Transcripts, outcomes, the "I'm not sure" moments. Two weeks of reading those is how an agent gets good.

Should you wait for GA?

For a business phone line that generates revenue when answered, we wouldn't run a preview service as the primary path. We would run it as overflow or after-hours on a secondary number, which is how we've soft-launched every receptionist anyway. The staged rollout in our 30-day plan (overflow first, then trusted call categories, then the main number) fits a preview product well.

Platform choice matters less than people think. Speech and telephony providers change their models and prices on a cadence of months; the booking logic is the thing the client paid for and the part that outlives any of them. Build so the Foundry piece, or the Twilio piece, or whatever replaces them next year, is a module.

Where this fits for a Microsoft shop

If your business runs on Microsoft 365 and Teams Phone, Foundry voice agents are the first option that keeps the whole call inside your tenant, with your compliance settings and your admin controls. That's a meaningful advantage for a law firm or a medical-adjacent practice that has to answer questions about where call audio goes. For everyone else, it's one more good option.

Our AI receptionist pricing post covers what the build and run costs look like across approaches. The telephony line in that budget just got smaller. The rest of it didn't.