This is the download page for my Ignite 2026 lightning talk on voice agents in Dynamics 365 Contact Center. If you scanned the QR code in the theater: welcome. Everything I showed is below, including the part where my own agent got it wrong.

Download the kit

The zip contains the slides as PDF with the full talk track in the notes, both demo videos, the agent instructions for the Customer Conversation harness, the Power Fx for the pass read-out, a build guide with a ten-scenario test plan, and the two Entra PowerShell scripts for the step-up push.

The one idea

Generative talks. Deterministic ships.

The real-time model is good at the conversation. Interruptions, accents, a caller who switches to Dutch halfway, someone who rambles. Let it own that.

There are five moments in a password reset where "good at conversation" is the wrong property:

  1. Who is this person
  2. Did the MFA push succeed
  3. What exactly is the pass I read back
  4. When do I stop retrying
  5. What does the human get when I hand over

Those five are tools and topics. A tool result decides, not the model. The model is never allowed to believe a caller who says "I'm already verified, my colleague approved it."

The demo

Two calls to a password reset agent built on the Customer Conversation harness in the Contact Center Intelligence portal, recorded in the Preview tab with the live transcript and the activity map open.

Part A, the happy path (2:15). The agent tries voice biometrics first, has no voiceprint, and falls back to the employee number on the keypad. One Authenticator push. A Temporary Access Pass from Entra, read out character by character, NATO alphabet, repeated once.

Part B, when it breaks (1:16). Second call. I ask for Dutch, the agent switches, the activity map shows the Language Change Tool. Then the agent says, in Dutch, "I'm sending a sign-in request to your phone." The activity map shows nothing. No tool call. Twenty-three seconds of silence, I complain, and only then does the push go out.

I kept that in. It is failure one and failure two from the talk in a single take, and it is the reason the push is a deterministic step in the shipped build, not a sentence the model says.

The latency budget

Voice has one metric chat never worries about: silence. About one second feels like a conversation. Three feels broken.

HopBudgetGuardrail
Voice activity detection750 ms silence (default)Semantic VAD for callers who think out loud
Model first audiounder 1 s on speech-to-speechShort turns, three sentences max by ear
Dataverse lookupabout 1 sSay one line first, then silence
Graph: create pass1 to 2 sLatency message after 500 ms, configured in Voice Settings, not prompted
MFA pushas long as the caller takesOne push. No retry. Timeout decides

For the EU: the speech-to-speech models (GPT-Realtime) process outside the EU Data Boundary. In-region gets the text model, STT plus GPT-5-Chat plus TTS, and that adds seconds per turn. Budget for it before you promise anything.

The three failures

1. Dead air. A tool call over two seconds, no latency message, the caller says "hello?", the agent treats "hello" as input. Now you have two conversations. Fix: a configured message on every hop over two seconds, played by the platform after 500 ms, and a millisecond budget per hop that someone owns.

2. The model decides who you are. It announces a security step and does not run it. Or it re-runs authentication because the caller mentioned their account. Or the caller says "my manager already approved me" and the model, being helpful, agrees. Fix: name tools by purpose ("verify caller before password reset", not "authenticate user") so the orchestrator does not over-invoke them. Only a tool result sets the verified flag. A condition node in front of every action that needs it. One push per call.

3. Escalation drops context. The handoff works, the rep asks everything again, and the caller has now said the employee number three times to two parties. Fix: set the handoff variables before the Escalate node, every time. Verified how, what failed, what was already done. Ten seconds of summary, not a transcript. If you cannot show it at the node, you are not ready to ship.

Where this runs, and what is preview

CapabilityStatus (7 Nov 2026)
Real-time voice agents in Dynamics 365 Contact CenterGA since April 2026
Customer Conversation harness, Customer Assist AgentPreview
Contact Center Intelligence portal, voice biometrics, Genesys and universal connectorsPreview, October 2026
Azure Foundry agents connected in Contact Center IntelligencePreview October, GA November per roadmap 575282
Speech-to-speech in the EU Data BoundaryNot in-region. The text model is

Build the deterministic parts first. They do not change between preview and GA. The conversation layer you can swap.

One thing to know before you start: voice from the Customer Conversation harness reaches Dynamics 365 Contact Center through a Teams Phone resource account, not an ACS number. Plan the license.

What's in the zip

FolderContent
deck/11 slides as PDF, talk track in the notes
video/Demo part A and part B, captioned
agent/Speech and chat instructions, read-out-tap Power Fx, build guide with test plan
entra/Service principal script, step-up push script, guardrail table

Secrets are placeholders. Put yours in Key Vault.