Status: Public Preview — behaviour may change.

What it does

Lets a representative pull a voice-enabled AI agent into a live voice call using the existing consult option. The agent runs a Copilot Studio topic — a mandatory disclosure, a consent capture, a PIN-based identity check — and then hands the call back. It all stays inside the same conversation.

Key facts

• The consult path is the only way a voice-enabled AI agent can join a call. There's no transfer or barge equivalent.
• Turned off by default. Enable it in Copilot Service admin center under Channels > Consult and transfer > Manage, then add agents individually under Support experience > Workspaces > Voice call experiences.
• Two per-agent settings you have to choose deliberately: whether recording and transcription pause during the consult, and whether the representative goes on hold.
• The agent needs a Transfer Conversation node to end its turn. Without one the consult times out.
• The AI agent must already be connected to the Voice channel and showing Connected in admin centre — a published Copilot Studio agent alone isn't enough.
• Prerequisites: Omnichannel administrator role and a configured voice channel.
• If the AI agent is the only consultee, the representative can't consult a human at the same time.

When to use / skip

Use it where a scripted, auditable exchange has to happen mid-call and you'd rather it wasn't the representative reading a script — regulatory disclosures, explicit consent capture, PIN verification. The pause-recording toggle is the point: it's a clean way to keep card numbers and PINs out of the recording without asking a representative to remember to hit pause.

Skip it for anything conversational. This is a workflow runner on a phone line, not a co-pilot. If the customer needs to actually discuss something, it'll go badly.

Configuration decisions

• Pause vs continue recording. Pausing protects sensitive data but leaves a hole in the audit trail. Decide which one your compliance team cares about more, and write it down — you'll be asked.
• Hold vs stay on the line. Keeping the representative on hold stops them talking over the agent, but they lose all context on what the customer said. For consent capture, hold. For a disclosure they'll need to reference afterwards, don't.
• One agent or several. Settings are per-agent, so a consent agent and an ID-verification agent can have different recording behaviour. Usually worth splitting rather than building one agent with branching.
• Timeout behaviour. Decide what the representative does when the agent fails — the system plays them a message, but your process has to say what happens next.

Gotchas

• The hold, transcription and recording settings apply to all consultees, not just the AI agent. Add a human consultee to a call that already has an agent configured to pause recording, and that human conversation isn't recorded either. Nobody expects this.
• Because the AI agent joins through consult, it shows up in consult reporting alongside human consults. Your supervisor metrics on consult volume will move, and it won't be because representatives started asking for help more.
• Preview banner means preview licensing terms. Check before putting it in front of a regulated workload, which is exactly the workload it's designed for.
• The Start topic runs automatically when the agent joins. If you've reused an agent that already does something on Start, it'll do that thing on the call.

Consultant notes

• Clients hear "AI agent joins the call" and scope a conversational IVR-in-consult. It isn't that. Set expectations in design: one deterministic workflow, then it leaves. The scoping conversation you want is "which script are we taking off the representative", not "what can it talk about".
• UAT surprise is almost always the recording gap. QA teams pull a sample call, find a silent stretch, and raise it as a defect. Brief them and the compliance owner before UAT starts, not after.
• Operationally, the dependency chain is longer than it looks: Copilot Studio agent, Voice channel connection, admin centre registration, workstream config. Any one of them drifting during an environment copy breaks the consult silently — the tab just doesn't show the agent. Add it to your post-deployment smoke test.
• Sequence this after the voice channel is stable. Teams that build it alongside an initial voice go-live end up debugging two things at once and can't tell which is which.

---

Preview feature — worth revisiting once it reaches general availability, particularly the recording and hold behaviour.