Microsoft Teams Voice Agents and Unify Integration, What is Possible?

Prefer audio?

Who should listen: Teams voice and contact centre owners, UC architects and M365 service owners looking at Teams Phone Agent, or weighing a native voice agent against a certified partner solution.

Guest: Gidi Adlersberg — Voca CIC Business Line Manager, AudioCodes

Teams Phone Agent is Microsoft's voice AI attendant, currently in the Frontier Program, and it was built to be extended from day one. AudioCodes is the first vendor certified under the new Microsoft Teams Voice Agent Certification Program. Gidi Adlersberg walks through what sits in the box, what the Unify integration model actually rests on, where certified partners add value, and the places voice agent projects tend to come unstuck.

Many thanks to AudioCodes, sponsor of this episode.

Key insights

• "AI receptionist" has become the industry shorthand for entry-level voice AI — every UCaaS vendor uses it, AudioCodes adopted it for Voca CIC, and Microsoft's flavour is Teams voice agents. Entry-level does not mean thin: Adlersberg reckons it covers a good sixty to seventy percent of the reasons people ring a main line. ▶ 4:21
• The native and the extended versions differ mainly in who looks after it. In the Microsoft world it is a licence activation — the Teams admin configures it, and when something goes wrong the Teams admin fixes it. The partner opportunity is the same capability as a fully managed service, covering planning, implementation and ongoing optimisation. ▶ 5:27
• Microsoft has placed the Teams Voice Agent Certification Program under Unify, the same integration model as the Teams Contact Center certification (the docs currently call it the Teams Unify Integration Model for Voice Agents). The message: whether you are building a contact centre or a voice agent on Teams, you use the same Azure stack. ▶ 8:07
• Unify rests on two Azure elements. The Calling SDK makes a vendor's application behave like a native Teams client for voice handling — no hairpinning, no API bridging. Call Automation orchestrates the call itself, and the voice agent program requires it: getting the call from the bot to a Teams user, the transfer, and the bot's audio streaming all run through it. ▶ 9:37
• You do not need the contact centre to get this. Voice agents come either as an add-on to Voca CIC agent licences or as a standalone buy — a main-line scenario with Teams Phone Agent on the front and AudioCodes building and managing the agent, with no contact centre underneath. ▶ 11:42
• The certification testing simulated an internet service provider support desk, with the bot working through why a caller's internet is slow — a scenario that leans heavily on harvesting a knowledge base to answer intelligently rather than reading from a script. ▶ 15:54
• The certification criteria went past the obvious. Alongside latency and accuracy of information, Microsoft tested whether the bot adapts in real time to caller sentiment and whether it holds a consistent level of empathy — the difference between walking a non-technical caller through finding the router and giving a technical one the short version. ▶ 16:56
• The experience is speech-to-speech, not speech-to-text and back again, which is where a lot of the latency in older bots came from. Under the managed service AudioCodes picks the underlying model, though customers can bring their own Azure AI Foundry model if they want to. ▶ 17:59
• Both agree the hard part is not the technology. More than half the work is scoping, the ROI case and getting business knowledge into the agent — and knowing how you will measure it. Defining what counts as a resolved case is less trivial than it sounds, and plenty of customers struggle with it. One organisation's voice AI looked broken in testing until they realised the gap was institutional knowledge their human agents carried that had never made it into the knowledge base. ▶ 21:30
• Escalation is a big part of the program, and Adlersberg calls poor escalation the modern disguise of the old IVR tree: instead of being stuck in press 1, press 2, callers get stuck with a bot that insists it can help. Certification covers escalating from bot to a Teams user based on presence and carrying the interaction context across, so the caller is not asked everything again. ▶ 23:32
• AudioCodes' commercial model is capped and deliberately free of token-based charges — priced per concurrent session and per agent, so the customer knows what the bill looks like at month end. Adlersberg likens where the market is heading to mobile in the nineties: bundles of minutes and texts first, fixed prices once it commoditised. ▶ 26:07

Insights summarised by AI from the episode transcript, reviewed by the Empowering.Cloud team.

Listen: Apple Podcasts · Spotify · Other platforms