Voice AI agents for customer service: the models are ready, so the handoff is what matters now

Our earlier piece on AI agents for customer service and quoting focused on chat and back-office workflows. This week's news is about the other channel, the one many organisations have been most nervous to automate: the phone. A few announcements, taken together, suggest that voice is the next place agentic AI becomes a normal part of service operations, and that the technology is no longer the hardest part.

The voice stack keeps getting faster

On 1 October, Microsoft announced three new models for voice applications through Microsoft Foundry: MAI-Transcribe-2-Streaming, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. The transcription model converts live speech to text as it is spoken, across 60 languages with automatic language detection. The two voice models generate speech, with the Flash variant aimed at situations where response time and call volume matter most. Microsoft positions them for customer service agents and multilingual assistants.

Treat vendor benchmark claims with the usual caution. Microsoft's transcription model is reported to rank first on an Artificial Analysis streaming benchmark, at a 2.5% word error rate. Benchmarks are not your call recordings, your accents or your product names. What is worth noting is the direction: transcription that starts producing text almost immediately is what lets an agent begin working out what a caller wants before they finish the sentence, and that is much of the difference between a conversation and an IVR menu.

This builds on Microsoft's April release of real-time voice agents in Copilot Studio, generally available with Dynamics 365 Contact Center in North America. Those agents are described as handling interruptions and moving from intent to action to confirmation in a single interaction, on tasks like billing, appointments and account changes.

Big brands are putting voice agents on the phones

The Home Depot has begun using an AI phone agent for in-store customer service calls, with plans to extend it to all US stores over the coming year, according to Customer Experience Dive. We would not read too much into any single retailer, but it matters that this is a high-volume, low-glamour use case: "is the product in stock, when does the store close, can someone help me". These are the calls where speed counts and the answer sits in a system somewhere.

What customers are telling us, via Gartner

Gartner's August survey is a useful counterweight to the enthusiasm. About half of customers say interactions are easier when companies use generative AI, but 87% say it is essential that a company offers a way to reach a human. Separately, Gartner reports that AI spending by customer service leaders has risen 38% while overall service budgets grew only 2%, which means AI is being funded largely by squeezing other lines. That raises the stakes on getting returns from it.

Our read: design the handoff first

Putting those threads together, we'd argue the competitive question has shifted from "can an AI hold a decent phone conversation" to "what happens when it shouldn't". A few practical takeaways:

• Start with narrow, verifiable calls. Status checks, bookings, simple account changes and after-hours coverage are better first candidates than complaints or cancellations.
• Make escalation a feature, not a failure. A good handoff passes the caller's intent, history and what has been tried to the human, so nobody repeats themselves. Microsoft's own materials make this point, and it is where many deployments will be judged.
• Be upfront that it is AI. Customers are already wary, and clear disclosure is cheaper than rebuilding trust.
• Connect it to your system of record. A voice agent that cannot see the customer's order, case or contract is just a more articulate dead end. For Salesforce customers, that means thinking hard about data quality and permissions before the voice layer, not after.
• Test with your own audio. Run real calls, with your accents, background noise and product jargon, before trusting a headline accuracy figure.
• Measure outcomes, not containment. Resolution, repeat contact and customer satisfaction tell you more than the percentage of calls the AI "handled".

As an independent advisor, this is where we tend to help: choosing between the platform options without a product to sell, designing the guardrails and escalation paths, and checking the risk side, from call recording and privacy obligations to what an agent is permitted to do on a customer's behalf.

The technology is ready for more of this work than it was a year ago. The organisations that do well will be the ones that treat the human handoff as part of the design from day one.

Sources: Microsoft Foundry and AI voice models coverage, PYMNTS, "Microsoft Targets Contact Center Market With Real-Time AI Voice Models"; Microsoft, "Build expressive voice experiences with new MAI models in Microsoft Foundry"; Microsoft Copilot Studio, "Extend AI voice support with real-time voice agents"; Microsoft Dynamics 365, "Dynamics 365 Contact Center AI Agents Transform CX"; Customer Experience Dive, "Home Depot adds AI phone agents"; Gartner, "87% of customers say companies using GenAI for customer service must provide access to a human agent"; Gartner, "AI spending by customer service leaders has surged by 38 percent".