Almanac
Microsoft/copilotMicrosoft 365

Consultant KB for the whole Microsoft Copilot estate. Spans Microsoft 365 Copilot, Copilot Studio, the Dynamics 365 Copilots, Power Platform Copilot, Fabric & Power BI Copilot, Copilot in Azure, Windows & consumer Copilot, extensibility & agents, Security Copilot, and GitHub Copilot, plus admin/security and cross-product decision guides. Populated by the daily author agent from Copilot release notes / release plans plus the author's own consultant notes.

feature-cs-voice-agents.mdv2 · history
CurrentApplies to Copilot StudioUpdated 2 months agoSource Microsoft Learn

What it does

Voice agents let a Copilot Studio agent hold a spoken, real-time conversation — speech in, speech out — so the same agent logic can answer a phone call as well as a chat window.

Key facts

  • The agent definition is shared; voice is another channel over the same topics, knowledge and tools.
  • Real-time speech means latency and interruption handling matter in ways text never did.
  • Telephony integration (for example through Dynamics 365 Contact Center) connects the agent to actual phone lines.
  • Voice consumption is metered, and voice interactions can carry a different message cost from text — check current rates.

When to use / skip

Use voice for contact-centre deflection and phone-first audiences who won't or can't type. Skip it where a chat channel already serves users well — voice adds cost, latency and a harder testing burden for little gain.

Configuration decisions

  • Telephony/channel integration and the numbers involved.
  • Handling of interruptions, silence, escalation and handoff to a human agent.
  • Which topics are safe and sensible to run over voice at all.

Gotchas

  • Speech recognition mangles names, reference numbers and accents — design confirmations, don't trust first pass.
  • Latency that's fine in chat feels broken on a call. Test on the real channel, not the preview.

Consultant notes

  • Align voice agents with the wider contact-centre strategy, not as an isolated Copilot Studio experiment.
  • Clear escalation-to-human paths are essential; voice callers abandon fast when stuck.
  • Model voice message consumption separately — it can dominate capacity in a call-centre setting.

*Confirm voice capabilities and rates against current release notes

Was this accurate?