Microsoft announced on November 18, 2026 that voice agents in Microsoft Foundry Agent Service are generally available. Microsoft describes it as "one enterprise-grade platform to build, deploy, operate and continuously improve" speech-to-speech agents, with built-in SIP support and connections to Teams Phone, Twilio and Genesys. We read the announcement on November 21, 2026. Every claim below links to it, and we mark which numbers come from Microsoft's own testing.
Key Takeaways
- Microsoft made voice agents in Foundry Agent Service generally available on November 18, 2026, with built-in SIP support for existing telephony and contact center systems.
- Microsoft's announcement lists Teams Phone, Twilio and Genesys as supported connections, and more than 40 industry-specific agent templates.
- Microsoft's 70.9 versus 63.7 score for Realtime Flash 1 comes from Microsoft's internal testing, and that model is still in preview.
- Foundry gives developers tracing and evaluation tools, so a team still builds, tunes and operates its own agent.
- Topcalls AI calls start from $0.25 a minute, depending on the country called, and a dedicated team configures each agent for the customer.
What did Microsoft ship for Foundry voice agents?
Microsoft shipped a full lifecycle platform for speech-to-speech agents. The post groups it into four parts: build and connect, observe and improve, connect to existing channels, and more natural voices. Developers create an agent, link it to company knowledge through Foundry IQ, add tools through MCP or function calling, and apply guardrails.
Observability is the notable piece. Microsoft says built-in voice-native tracing lets a developer inspect a call turn by turn and see whether a slow response came from speech recognition, the model, a billing-system tool call or speech synthesis. Teams can replay the conversation when transcript and audio capture are on.

The announcement also says that "thousands of customers" adopted voice agents since the public preview. It gives no list of those customers beyond three quoted ones, and no revenue or call volume.
Which phone systems do Foundry voice agents connect to?
Foundry voice agents connect to the web, to Teams Phone, to Twilio and to Genesys. Microsoft says built-in Session Initiation Protocol (SIP) support lets agents "connect directly with existing contact center and telephony infrastructure," so a company can keep its current phone setup.
This is a real strength for a business already inside Microsoft's stack. A customer who lives in Teams Phone gets an agent that hands a caller to a person in Teams on the same call. Landis Technologies' CEO, Matt Landis, says in Microsoft's post that after a regional bank deployment, "our AI agent now handles 30% of calls from end to end." That is a partner's statement about one customer, so it shows what is possible and not what to expect.
Microsoft's post is about inbound and contact center routing. It does not describe an outbound campaign dialer, with lead lists, retries and call windows, so check that point if outbound volume is your goal.
Are Microsoft's benchmark numbers independent?
No. The two performance figures in the post come from Microsoft's own testing. Microsoft says Realtime Flash 1 scored 70.9 on the τ³-Voice benchmark in "internal testing," against 63.7 for GPT-realtime-2.1, with gains in telecom and retail and a six-point drop in airline.
The second figure covers turn-taking. Microsoft reports that on Full-Duplex-Bench v1.5, a public benchmark, its azure_semantic_vad v2 raised the total score from 46.3% to 72.2%, a 25.9-point gain. The benchmark is public, but the run is Microsoft's.
| Claim | Who measured it | Status |
|---|---|---|
| Realtime Flash 1: 70.9 vs 63.7 on τ³-Voice | Microsoft, internal testing | Model in preview |
| Turn-taking: 46.3% to 72.2% on Full-Duplex-Bench v1.5 | Microsoft, public benchmark | Generally available |
| Agent handles 30% of calls end to end | Landis Technologies, one bank customer | Partner statement |

Treat each as a vendor claim until you hear your own calls on it.
Which models can power a Foundry voice agent?
Microsoft lists GPT Live 1 and Microsoft Realtime Flash 1 next to its existing GPT Realtime models. GPT Live 1 is a full-duplex model that can hand tasks to Foundry prompt agents and hosted agents. Realtime Flash 1 aims at low-cost serving for customer support, appointment scheduling and routine service requests, and Microsoft says it is "now in preview" at Voice Live Lite tier pricing.
Microsoft does not print a per-minute price in the post, so we cannot compare cost. Ask for a total per call that includes speech, the model and telephony. We cover how OpenAI's side of this stack works in our GPT-Live API guide.
Languages are a second check. Microsoft says real-time transcription now recognizes 50+ languages natively, and its new HD V3 voices add English, Chinese, Filipino, Indonesian, Malay and Indic languages. Topcalls agents call in up to 78 languages, and how many you get depends on the transcription and voice models you choose. See multi-language support.
Should you build on Foundry or buy a managed service?
Build on Foundry if you have developers, you already run Microsoft telephony, and you want to own the agent. Buy a managed service if you want calls running without building the agent, the tracing and the telephony yourself. Foundry is a platform, and Microsoft's own wording is "build, deploy, operate."
Four questions separate the two paths:
- Who writes and tunes the prompts? On Foundry your team does. At Topcalls a dedicated team configures everything, so you need no prompt engineer.
- Who watches quality? Foundry gives you traces to inspect. Topcalls returns a recording, a transcript, a summary, a disposition and a post-call analysis for every call.
- Who owns outbound? Foundry's post covers contact center channels. For campaigns with retries, see smart campaigns.
- What is the total per call? Topcalls AI calls start from $0.25 a minute, depending on the country you call, billed by the second, with the AI voice model, call recording, transcription and analytics included.
Our build vs buy guide puts numbers on this choice, and the voice agent platform guide compares the categories. For the product itself, see AI voice agents.
To price your own volume, run the ROI calculator or talk to us.
Sources
- Voice agents in Microsoft Foundry Agent Service now generally available, Microsoft Foundry Blog, November 18, 2026.
Frequently Asked Questions
Get AI calling tips in your inbox
No spam. One email per week with actionable sales automation tips.






