14 August 2026
An AI agent in a meeting today can reliably answer questions from past transcripts, draft follow-up messages from what was said, and push structured updates into a connected tool once someone has wired that connection. It cannot reliably make a judgment call on your behalf, negotiate a decision, or file work without a human checking it first. The gap between those two lists is most of what's being oversold right now.
This is the part that holds up under actual use. When someone asks 'what did we agree on pricing last Tuesday,' the agent is searching a timestamped transcript and returning the relevant span, sometimes with a summary layer on top. That's a well-bounded problem: the answer already exists in text, the agent just has to find it and phrase it back. It fails in predictable ways — if the transcript is garbled because someone was on a bad connection, or if the meeting used a platform where speech recognition never ran cleanly, the answer will be wrong or incomplete, and it won't tell you that.
AVAY's version of this lets someone address the AI participant directly during a live call and get an answer pulled from any past meeting the team has had, spoken back out loud, not just typed into a chat panel. The same limitation applies as everywhere else: transcription quality depends on browser speech recognition, so accuracy is meaningfully better in Chrome and Edge than elsewhere.
Agents are good at producing a first draft of a recap email or a Slack update from what was said in a call — who owns what, what was decided, what's still open. The mechanism is straightforward: it's summarizing structured notes it already kept current during the meeting. Where it stops being reliable is tone, audience, and what to leave out. An agent doesn't know that the client on the call gets defensive about deadline language, or that a decision needs to be softened before it goes to a stakeholder who wasn't in the room. That judgment isn't built into these systems, and nobody selling this feature claims it is when pressed — but the marketing copy around 'automated follow-ups' often implies more autonomy than exists.
In practice this looks like: draft appears within a minute or two of the call ending, a person reads it, edits two or three lines, and sends it themselves. The time saved is real — it's the difference between writing from scratch and editing — but it is not zero human time, and treating it as such is where teams get burned.
This is the capability most exaggerated in demos. An agent creating a Jira ticket or updating a CRM field mid-call looks impressive, but it only happens because someone configured that specific connector to that specific tool, with specific permissions, before the meeting started. There's no general-purpose ability to 'reach into your systems' — there's a set of integrations someone built, and the agent operates inside whatever those integrations expose.
AVAY's connectors work this way: a team attaches them to their own systems, and the AI can act through what's attached — nothing more. If there's no connector to your ticketing system, the agent can note the action item in the meeting record, but it cannot file the ticket. That's a real constraint, not a caveat buried in the fine print, and it's worth checking before assuming an agent will 'just handle' downstream work.
A few phrases are doing a lot of work in vendor pages right now, and it's worth naming them plainly. 'The AI runs the meeting' generally means it keeps notes current and can answer questions when addressed — it is not facilitating discussion, redirecting off-topic conversation, or making a call on which of two proposals to pursue. 'Autonomous task execution' usually means a draft or a suggested action sitting in a queue for approval, not something that ships without a person clicking approve. 'Understands context across your whole org' typically means it can search meetings it was present for or transcripts it was given — not that it has any awareness beyond what's in the room.
Every genuinely useful part of this — the recap, the ticket, the CRM update, the answer to a question — passes through a moment where a person decides whether it's right before it goes anywhere consequential. That review step is not overhead being eliminated by AI; it's the thing that makes the AI's output safe to use. The teams that get burned are usually the ones that skip it because the draft looked polished enough to trust, and the ones that get real value are the ones that treat the agent as producing a fast first pass, not a finished decision.
| What's marketed | What actually happens | |
|---|---|---|
| Answering questions | "Knows everything the team has discussed" | Retrieves and quotes from transcripts it has access to |
| Follow-ups | "Automated follow-up emails" | Produces a draft; a person edits and sends it |
| Filing work | "Acts across your tools" | Acts only through connectors someone has attached |
| Decisions | "AI-driven decision making" | Surfaces what was said; the decision stays human |
No current product does this in a way anyone should trust for real decisions. Agents can transcribe, take notes, and answer questions when a human is present and addresses them, but there's no version of this that facilitates a meeting, resolves disagreement, or represents a party's interests unattended.
It can draft the action — a ticket, a message, a calendar hold — but completing it in a way that affects another system almost always requires either a connector someone configured in advance or a human clicking to confirm. Treat "completed" claims in a demo as "drafted and ready for approval" until you've verified otherwise.
Not exactly. Transcription quality in-browser depends on the browser's speech recognition engine, which is why accuracy is generally better in Chrome and Edge. Platforms that require a download or a bot to join a third-party call add another layer of latency and failure points that browser-native tools don't have.
It should say it doesn't know, but not all of them do that cleanly — some will produce a plausible-sounding answer stitched from adjacent context, which is worse than no answer. Test this deliberately with a question you know it can't answer before relying on it for questions you can't verify yourself.
It's a trade-off, not a clear win. Connectors you attach yourself mean the agent only touches systems you've explicitly authorized, which is safer but means nothing works until you set it up. A vendor with pre-built integrations to major tools works out of the box but gives you less control over exactly what it can see and touch.
An AI agent in a meeting is reliable at retrieval and drafting and useless at judgment — the honest way to use one is to let it do the recall and the first draft, and keep a human in the loop for anything that leaves the room.
Meetings that take their own notes, in the browser: avay.ai.