Real-Time Voice Agents (Speech-to-Speech IVR)

What it does

Real-time voice agents let you build speech-to-speech IVR experiences in Copilot Studio using OpenAI's gpt-4o-realtime model via Microsoft Foundry. The caller talks naturally, the AI responds in real-time audio — no DTMF menus, no turn-by-turn text intermediary. GA from April 2026.

Key facts

Built and deployed entirely within the Copilot Studio authoring experience — same tooling as standard voice agents
Powered by OpenAI gpt-4o-realtime through Microsoft Foundry (Azure-hosted); data stays within the Microsoft trust boundary
Designed for low-latency, natural voice interactions — Microsoft targets responsiveness comparable to human conversation
Supports core IVR scenarios: intent capture, routing decisions, simple self-service tasks
Separate from standard Copilot Studio IVR voice agents (which use TTS/STT with a text reasoning layer); this uses a native audio-to-audio model
No DTMF fallback baked in by default — you design the interaction model

When to use / skip

If your client's IVR containment rate is bottlenecked by unnatural, stilted bot conversations, this is worth evaluating — the speech-to-speech model handles interruptions and conversational repair better than the standard STT/TTS pipeline. Skip it if your client's IVR is handling structured, menu-driven calls well and you don't have a clear containment problem to solve. Also skip if the client is in an early routing phase — get the routing and basic self-service working first, then layer this in.

Configuration decisions

Which IVR scenarios to target with real-time voice vs. standard voice agents — don't migrate everything at once
What happens at handoff: real-time voice agents still escalate to the queue, so your routing configuration doesn't change materially
Foundry model access: confirm the client's Azure subscription and region support gpt-4o-realtime through Foundry
Latency tolerance: the model is low-latency but not zero-latency — check whether your client's call quality SLA accounts for the added inference time

Gotchas

This is a distinct authoring experience within Copilot Studio, not a toggle on existing voice agent flows — you're building new bots, not upgrading old ones
"Iterate and optimize continuously" in the Microsoft docs means you can update the agent without a full republish cycle, but it doesn't mean zero-downtime deploys — test in a dev environment
Real-time voice uses substantially more AI credits per interaction than standard IVR — factor this into licensing conversations. The WFM AI credit estimation feature can help project consumption
Microsoft's geo-availability for Foundry gpt-realtime models is narrower than standard Azure OpenAI regions; check the Feature Geography report before committing to clients in non-US regions

Consultant notes

Clients will want to use this for everything once they see a demo. Frame it early: real-time voice agents excel at conversational self-service (e.g., account queries, booking changes) but add cost and complexity that's not justified for simple IVR deflection of "press 1 for billing"
Watch the credit consumption angle carefully — this will catch clients out at renewal if the design is sloppy. Get an estimate from WFM forecasting before signing off on the design
The Copilot Studio authoring experience is still the same canvas as before; what changes is the underlying model. If your team is comfortable building standard voice agents, the learning curve is lower than it looks
Post-go-live, you'll want to monitor latency and abandonment separately for real-time voice agent queues — the user experience failure mode is different (callers hanging up when the AI pauses to think) rather than the standard bot failure mode (misunderstood intents)

Worth revisiting as Microsoft expands the scenario coverage and region availability through the rest of 2026 wave 1