What Is an AI Meeting API?

An AI meeting API lets you programmatically create video meetings that an AI participant runs, it joins, sees, hears, speaks, pursues goals, and returns a transcript, recording, and structured analysis when the call ends. You call one endpoint to get a scenario, another to create a meeting, and a real AI persona holds the conversation: an interview, a sales roleplay, a discovery call, a testimonial ask. No human on your side needs to be in the room.

Anyone building a product that used to require a person on a video call, screening candidates, training reps, gathering requirements, collecting feedback, is the audience for this. The reason it's a category now and not five years ago: realtime voice models got good enough to hold a real conversation, and vendors started wrapping the video room, the turn-taking, and the scoring around them so you don't have to build all three yourself.

What an AI meeting API does

Strip away the marketing and an AI meeting API gives you a fixed set of primitives, called over HTTP:

• Create scenarios and personas, define who the AI is (name, job title, demeanor, background) and what it's supposed to accomplish in the call.
• Spin up meeting sessions, one API call returns a live video room URL a participant can join immediately, no scheduling required.
• Real-time voice and video conversation, the AI persona talks, in turn, for the duration of the call, not a scripted IVR tree.
• Vision, camera and screenshare input the AI can see and reason about mid-conversation, not just hear.
• Evaluation goals and scoring, criteria you define up front that get scored against the transcript once the call ends.
• Transcripts, recordings, and structured analyses, pulled from a single endpoint or pushed to you the moment they're ready.
• Webhooks, event notifications for every stage of the session lifecycle, so your backend doesn't have to poll.

That's the shape. The differences between vendors are in how well each of these seven things is built, not whether they exist at all.

How it differs from adjacent APIs

"AI meeting API" gets used loosely. It's worth being precise, because the four categories people confuse it with solve genuinely different problems.

The distinction that matters most in practice: a meeting-bot API assumes two humans are already talking and just watches. A voice agent API assumes there's no video surface at all. A speech-to-speech model API assumes you're building everything, including the parts that make a conversation feel like a meeting rather than a chatbot with a microphone. An avatar API assumes the win is a convincing face, and worries less about whether the thing behind the face holds a personality for thirty minutes without drifting.

An AI meeting API is the only one of the five that ships the video room, the AI-led conversation, and the post-call scoring as one connected thing. If you're deciding between building this yourself on a raw model or buying the category, that trade-off is its own post, see build vs. buy for an AI meetings API.

Core primitives

Every AI meeting API is built from a small vocabulary. Waterr's version, in its own docs, breaks down to six pieces:

• Scenario, the complete conversation package. A blueprint combining who the AI is, how it behaves, and what it measures. Everything starts here.
• Persona, the character your participant talks to. Name, job title, demeanor, industry, background, avatar. This is what shows up on screen.
• Meeting script, the instructions that drive behavior turn to turn: how to open, what to cover, how to handle specific situations, the overall flow.
• Goals, the evaluation criteria the AI scores the participant against after the session. Invisible to the participant during the call.
• Meeting session, one actual conversation instance. A participant joins a link, a session starts, and it produces a real-time conversation, a recording, a transcript, goal scores, and speech analytics.
• Analysis, the scored output: average score, per-goal feedback, strengths, growth areas, highlights tied to transcript timestamps.

Get these six right and the rest of the API is just CRUD around them.

How simple is the integration

Simpler than the capability list suggests. Every new Waterr account comes seeded with two ready-made scenarios, so the first live AI meeting takes exactly two API calls: list your scenarios, then create a meeting from one. The response hands back a meeting link your participant can open right away - in a browser, or embedded inside your own product. Once the session ends, one more call returns the summary, transcript, and goal scores together.

Building your own persona from scratch instead of using a seeded scenario adds one call to that flow. The full request-level walkthrough lives in the quickstart at docs.waterr.ai - hand it to your engineering team and the first working meeting is an afternoon, not a sprint.

If you're specifically building an AI-led candidate screen or interview, the primitives are identical but the goal design changes, see the AI interviewer API and how to build an AI interviewer for the goal and script patterns that work for hiring specifically.

What to look for when choosing one

Every AI meeting API will demo well in a sales call. The differences show up three weeks into integration. A few things worth checking before you commit:

Turn-taking quality. Does the AI cut people off mid-sentence, or wait an awkward beat too long? This is the single biggest tell for whether a vendor has done real engineering on the conversation layer versus wired a default voice-activity-detection config and shipped it.

Vision. Can the AI see the participant's camera and screen share mid-call, or is "vision" just a post-call transcript of what was said about the screen? Waterr's vision capability reads facial expressions, body language, and screen content live, reviewing code, reading a shared doc, reacting to a confused expression, and you don't have to instruct it to "use vision"; it incorporates visual context automatically when camera or screenshare is active.

Evaluation and scoring built in. Do you get goal-based scoring out of the box, or do you have to build your own LLM-as-judge pipeline on top of the raw transcript? Waterr scores each defined goal on a 1-10 scale with written feedback, tied to specific transcript moments.

Webhooks. Do you get signed, retried, per-event webhooks (meeting.ended, transcript.ready, session.analysis_complete), or are you stuck polling an endpoint on a timer? Signed webhooks with HMAC verification and durable retries are the difference between a production integration and a cron job that occasionally misses events.

Observability. Can you pipe STT/LLM/TTS spans and token usage into your own OpenTelemetry backend, Langfuse, Honeycomb, Datadog, or is the pipeline a black box you can only debug by staring at the transcript?

White-label. If you're shipping this inside your own product, can participants see your brand and domain instead of the vendor's, or are you stuck with someone else's logo on every session? Waterr's white-label option replaces branding on the session interface, share links, and emails, gated to enterprise plans, worth checking before you build a customer-facing flow around a vendor you can't rebrand.

Embed options. Inline, floating button, or corner avatar widget, dropped into a page with one iframe, or requiring a signed token for private scenarios. This matters if the AI meeting is a moment inside your product rather than a standalone link.

MCP / agent-skill support. Can your own coding agent or Claude Code list scenarios, pull meetings, and check analysis scores without you writing glue code? An MCP server means your agent calls the same API you'd call by hand.

If your use case is specifically avatar-heavy, you care more about the face than the conversation, the avatar-API comparison is a more direct read than this general category post.

Frequently asked questions

Is an AI meeting API the same as a meeting bot? No. A meeting bot joins an existing human-to-human call to record and transcribe it, it doesn't talk. An AI meeting API creates a session where the AI itself is a full participant: it converses, pursues goals, and drives the meeting.

Can an AI meeting API see screen share? With vision capability enabled, yes. The AI can read documents, review code, interpret charts, and reference specific things it sees on a shared screen mid-conversation, not just transcribe what was said about it afterward.

How do I get the results of an AI-run meeting? One API call once the session ends. You get a summary, full transcript, per-goal scores with feedback, strengths, and growth areas, typically ready 30-60 seconds after the call finishes. Webhooks can push this to you instead of polling.

Can I embed an AI meeting in my app? Yes, inline, as a floating button, or as a corner avatar widget. Public scenarios embed with a plain iframe snippet; private scenarios or per-user identity require a short-lived signed embed token minted server-side.

How is an AI meeting API different from a speech-to-speech model API? A speech-to-speech model is a raw bidirectional audio stream, you send audio, get audio back. It has no concept of a meeting room, a persona, a goal, or a scorecard. An AI meeting API is built on top of a model like that, adding the video room, the turn-taking engineering, the persona and scenario layer, and the post-call analysis, so you don't build all of it yourself.

The category is young enough that the vendor list is still short and the naming is still loose. That won't last. The teams shipping real products on this right now are the ones who picked apart what "AI meeting API" means before they wrote a line of integration code, and picked the primitives that matched what they were building, not the ones that demoed best.