Voice AI Agents in Meetings: What Actually Talks, and What It Costs

Corrected October 3, 2026, and rewritten October 4, 2026. An earlier version described hands-on tests of five voice AI platforms. No testing took place; meetingstack.io is a desk-research publication. This version uses only public vendor pages, platform documentation and one independent benchmark, each linked and checked October 4, 2026. See our corrections page.

AI agents that talk out loud in a meeting are real, and you can buy several today. They cost roughly 4 to 13 times what a silent Recall.ai notetaker bot costs per hour at list prices. Zoom steers developers to a listen-only API, and Google has closed its media API to new signups. And the only speaking agents Microsoft, Google and Zoom ship themselves are translators.

That is the short version. Below is the evidence, in four parts: who can speak, what the platforms allow, what an hour costs, and how fast these agents really answer.

Which agents can speak in a meeting today

Three kinds of product can put a voice into a Zoom, Teams or Google Meet call.

Notetakers that answer back. Otter turned its assistant into a "voice-activated AI meeting agent" in March 2025, starting on Zoom. Fireflies lets anyone say "Hey Fireflies" in a live call, and with the "Respond via voice" setting on, it replies "with audio instead of text in the meeting chat." Fireflies says the feature is "free to use". Both answer questions on request. Neither joins the discussion on its own.

Meeting-bot APIs with audio output. Recall.ai includes "display video and play audio in the meeting" in its $0.50 per hour bot price, and its AI Agent API can stream audio and video into six platforms. MeetStream sells voice agents that "join your Zoom, Google Meet or Microsoft Teams call as a participant", running either a single realtime model (OpenAI, Gemini or xAI) or a speech pipeline such as Deepgram, OpenAI and ElevenLabs. Attendee publishes an open-source demo that wires its bot to Deepgram's Voice Agent API. MeetGeek offers Voice Agents that run a scripted conversation from your instructions, in public beta.

Phone-agent platforms. Vapi, Retell, Bland and ElevenLabs build voice agents for phone and web calls. Their own pages describe phone and web calls, not meetings. To get one into a Zoom or Teams call, you pair it with a meeting bot from the group above.

The beta notes are where you learn what is hard. MeetGeek's own guide warns that "in some Zoom meetings, you must manually unmute the agent's microphone for it to speak," and that the host must admit the agent from the lobby. Those two lines describe the real obstacle: a speaking agent is a bot, and the platforms decide what bots can do.

What Microsoft, Google and Zoom ship themselves

The three platform owners all sell meeting agents now. Their general-purpose assistants do not speak.

• Microsoft Facilitator works by chat. Licensed users mention it in the meeting chat, and it posts its reply "within the meeting's conversation for everyone to see." It needs a Microsoft Copilot license and does not work in external meetings (Microsoft Learn, updated August 13, 2026).
• Ask Gemini in Google Meet is a private text assistant: "A user's interactions and responses from Ask Gemini in Meet are private to that user" (Google, September 17, 2025). Google extended it to Business Standard in January 2026.
• ZoomMate, which replaced the AI Companion brand on June 24, 2026, adds a custom agent builder, agentic search and workflows. Its agents act on meeting context; they do not take a turn in the conversation.

The agents that do speak are translators. Teams' Interpreter agent provides "real-time speech-to-speech interpretation" in 10 languages and can generate the translation "in your own voice"; a Copilot license includes "20 hours of interpretation ... per person per month" (Microsoft Support). Google made Meet speech translation generally available on January 27, 2026 and put a larger model, covering "over 70 languages," into private preview for Meet in June 2026. Zoom lists a Voice Translator for "live audio translation for Meetings" in its paid ZoomMate tier.

Our reading: translation is the one job where an AI voice speaking in a meeting has a fixed script (repeat what a human just said, in another language) and a clear owner (the person being translated). The platforms have shipped that and nothing more open-ended.

Will the platform let a speaking bot in?

A notetaker only needs to listen. A speaking agent needs to send audio into the call. In October 2026, only one of the three platforms offers a first-party API that does that. Teams and Meet now screen third-party bots by default; Zoom limits unpublished SDK bots to internal meetings and is steering developers away from bots.

Microsoft Teams has the formal path. Its Real-time Media Platform lets a bot "send and receive voice and video content frame by frame". But a policy that Microsoft announced on August 21, 2026 (MC1459141) sets the default for external bots to RequireApprovalWhenDetected, with general availability scheduled to finish by the end of September 2026 (Office 365 for IT Pros). Microsoft says detected bots need organizer approval "even when a meeting allows participants to bypass the lobby." Microsoft has worked with Maestro Labs, Otter.ai, Read AI and Recall.ai to register their bots (Office 365 for IT Pros), and the Bot Identification Program's public preview "has reached its current bot provider capacity" (Microsoft Learn). We covered the policy in Teams now holds recording bots in the lobby.

Google Meet has no supported way for a bot to talk. The Meet Media API lets apps "consume" audio, video and participant metadata, and it is "no longer accepting new signups". Speaking bots on Meet join through a browser like a human guest, and since March 2026 Meet sorts join requests into two queues; the queue for flagged joiners, which "could include automated bots," is "automatically denied unless the host intervenes" (UC Today, March 25, 2026).

Zoom points developers to Realtime Media Streams, which Zoom describes as "a data pipeline that gives your app access to live audio, video, and transcript data" instead of putting bots in meetings. RTMS only receives: Recall.ai's guide states "you cannot send media back into the meeting through RTMS" (Recall.ai). A speaking agent on Zoom therefore needs a meeting bot, typically built on the Meeting SDK, and on September 10, 2026 a Zoom staff member told developers, "We are defocusing on SDK use-case for AI notetaker / Bots. All apps publishing are subjected to approval process. The recommended path forward is to utilize RTMS" (Zoom developer forum).

Put together: the official listening APIs are growing, and the path a speaking agent depends on is the one under the most scrutiny. For the wider shift, see Recording without bots: the post-bot era.

What a speaking agent costs per meeting hour

Phone-agent platforms bill per connected minute, which suits a four-minute appointment call. A meeting runs an hour, and an agent that speaks for ten of those minutes is still connected for sixty. That billing mismatch drives the numbers below.

At these prices a speaking agent costs about 4 to 13 times a Recall.ai notetaker hour, and up to about 29 times at Retell's top rate. Meeting-bot pricing itself is covered in detail in what a meeting-bot hour really costs.

Three caveats keep this honest. First, we assumed the voice session bills for the whole hour. MeetStream advertises wake-word gating, and an agent that connects its voice stack only when addressed would cost far less. Second, AgentCall does not define the unit behind "voice intelligence"; we applied it to all 60 minutes. Third, pairing a voice platform with a meeting bot is an architecture MeetStream and Attendee document, but we did not confirm that every voice vendor supports that bridge on every plan.

For comparison, two speaking features cost nothing extra if you already pay for the host product: Talk to Fireflies is free, and Teams Interpreter includes 20 hours a month with Copilot, which Microsoft lists from $21 per user per month for Copilot Business (shown at $18 on promotion as of October 4).

Latency: vendor claims vs a caller-side measurement

Speed matters more in a meeting than on a phone call, because other people are waiting to talk. Every voice vendor advertises a fast response. One independent benchmark measures what a caller actually hears.

Measured figures come from the OpenBenchmarks voice agent latency board, which records real phone calls and times the gap from the caller going quiet to the agent's first sound. Its last call was placed August 1, 2026, with 382 to 429 usable turns per platform, and its code and per-call data are public. Telnyx posted the lowest median at 1,296 ms. OpenBenchmarks describes itself as independent; it does not name its operators, and its customers get validation sets.

The gap is not proof that vendors are lying. The two numbers measure different things. Vendors time the reply at their servers; the benchmark times the silence a caller hears, including the phone network and end-of-speech detection. Its method page says the latency a platform reports for itself "runs roughly 490 ms below what we measure from that call's audio." Retell's homepage says "independent benchmarks confirm Retell as the leader in responsiveness"; on this board it placed fifth of five.

Two limits apply. These are phone calls, and a meeting bot adds its own hop between the call and the agent. MeetStream claims "under 100 ms" for that hop and about 300 ms end to end in realtime mode, and says customers can check per-bot latency metrics themselves, but nobody has published an independent measurement of agents inside meetings that we could find.

Even Microsoft's own speaking agent comes with turn-taking advice. Its Interpreter guidance tells participants to "raise your hand ... preventing overlapping speech" and notes the agent is not optimized for rapid exchanges. If the platform owner asks humans to queue for a translator, expect a third-party agent in an open discussion to struggle with interruptions.

Where speaking agents make sense now

This section is our analysis of the evidence above, not test results.

The jobs that fit share three traits: someone asks the agent to speak, the answer is short, and one person owns the exchange.

• On-demand answers. "Hey Fireflies, what were the action items from the first 10 minutes?" is the example Fireflies itself gives. The agent speaks only when called.
• Translation. Fixed script, clear owner, and the platforms already ship it.
• Scripted one-to-one calls held on a meeting link. MeetGeek's templates for screening and qualification calls, and its Calendly setup where the agent joins booked Zoom calls, are phone-agent use cases moved into a video call.

Open multi-party discussion is the poor fit: several speakers, interruptions, references to a shared screen, and a 1.5-second gap after each turn at phone-call medians. No vendor we reviewed publishes data showing an agent handling that well.

How to evaluate a speaking meeting agent

1. Does it speak, or only chat? Facilitator, Ask Gemini and ZoomMate agents work in text. Check the vendor's own docs for audio output.
2. Which platform path does it use? A registered Teams bot, a Zoom Meeting SDK app, or a browser bot on Meet each carry different approval and lobby rules. Ask how the vendor handles Teams' ExternalBotAccessMode in your tenant.
3. What is the billing unit? Per connected minute or per speaking minute changes the voice cost about six-fold in our one-hour example.
4. Measure latency from the listener's side. Record both sides on one clock, as the OpenBenchmarks method describes, rather than trusting a dashboard figure.
5. Who admits it, and who is told? Teams and Meet put the decision in the organizer's hands by default; on Zoom, check how your meeting's own admission settings treat bots. Plan for the agent to wait in a lobby.

See all voice tools in our voice AI category, compare platforms in Vapi vs Retell and ElevenLabs vs Vapi, or start from the bot layer at meeting bot as a service.

FAQ

Can an AI agent speak in a Zoom meeting?
Yes, through a bot that joins as a participant, such as products built on Recall.ai, MeetStream or Attendee, or Otter's and Fireflies' voice features. Zoom's RTMS API cannot send audio, and Zoom staff said in September 2026 that Zoom is "defocusing" on SDK bots.

Does Microsoft Teams have an AI agent that talks in meetings?
The Interpreter agent speaks, translating in 10 languages, with 20 hours a month included in a Copilot license. Facilitator, Microsoft's meeting agent, replies in the meeting chat.

How much does a voice AI meeting agent cost?
At October 2026 list prices, about $2.76 to $8.23 per meeting hour for the setups we priced (more at Retell's top rate), against $0.45 to $0.65 for a silent notetaker bot. Costs drop if the voice session runs only when the agent is addressed.

How fast do voice AI agents respond?
Vendors claim 400 to 600 ms or "sub-second." An independent benchmark of phone calls measured medians of 1,296 to 1,740 ms across five platforms in August 2026. The difference is partly how each side measures.

Method

Desk research only. We did not test any product. Every figure comes from a public page linked inline, checked October 4, 2026: vendor pricing and product pages, Microsoft, Google and Zoom documentation, two third-party reports (UC Today, Office 365 for IT Pros), one vendor explainer (Recall.ai) and one independent benchmark (OpenBenchmarks, phone calls, August 2026). Cost figures are our arithmetic on list prices, with assumptions stated next to the table. We excluded figures we could not load from the source, including ZoomMate's per-user price and analyst forecasts on agent adoption. See our methodology.