Teams Now Holds AI Meeting Bots in the Lobby

If an AI assistant works by joining a call as a meeting participant, Microsoft Teams now detects it and holds it in the lobby until a human admits it. That is the default setting, not an option an administrator has to switch on, and it applies to meetings hosted by the organization that owns the Teams tenant — which, in a job interview, is the employer.

This matters in both directions. The employer's notetaking bot is subject to it, and so is any assistant a candidate sends into the call. Anything that gets its audio by becoming an attendee now depends on someone on the hiring side clicking Admit while looking at a warning about what bots do. Facts below are current as of 27 September 2026 and drawn from Microsoft's admin documentation and dated reporting on the rollout.

What Teams changed

Microsoft's admin documentation describes a meeting policy setting named "Manage external bots and their access to meetings," which appears under Meeting join and lobby and maps to the PowerShell attribute ExternalBotAccessMode. It has two values. AllowBots means bots "aren't detected or marked in the lobby or in the meeting. They appear as any other external participant in the meetings." RequireApprovalWhenDetected, which Microsoft documents as the default, means bots "are detected and marked in the lobby. They're enforced to be in the lobby until explicitly admitted into the meeting" (Microsoft Learn: Manage external bots and their access to meetings).

Two details in that documentation do more work than the headline. Detected bots are lobbied "irrespective of the lobby configuration for the meeting," so an interview set up to let everyone straight in still stops a bot. And Microsoft states that organizers "are reminded about the risk of admitting bots into their meeting to ensure bots aren't mistakenly admitted from the lobby" — the admission decision is framed to the recruiter as a risk, at the moment they make it.

Detection is inference rather than a registry lookup. Microsoft says Teams "uses built-in detection mechanisms to identify external AI bots based on signals collected during the meeting join process," combining "infrastructural and behavioral signals," and acknowledges that "some external bots may not be detected." Undetected bots appear as ordinary participants.

The rollout arrived in stages through 2026. Reporting in May 2026 described the first wave, announced in March, as detection and labeling of unverified bots (UC Today, 6 May 2026). The administrator policy followed: Office 365 for IT Pros dates the message center notice to 21 August 2026 and general availability as completing by the end of September 2026, and notes that Microsoft identifies registered independent-vendor bots such as Otter.ai as named apps while unregistered ones surface as suspected threats (Office 365 for IT Pros, 24 August 2026).

The policy belongs to whoever hosts the meeting

The setting applies to meetings hosted in the tenant that configured it. For an interview, that is almost always the employer's tenant, which means you cannot change the rule, cannot see which value is set, and will not find out until a bot either joins or sits in the lobby with a label on it.

That asymmetry cuts both ways for a candidate. It is the same structural point as the employer's own notetaker: a bot in the participant list is an attendee someone chose to admit, which is why reading the participant list before you speak is worth the three seconds. A proctoring or interviewer bot that joins a live call is the AI participant you were not introduced to — and now, on Teams, one that had to be waved through.

Zoom and Google Meet: the same chokepoint, different labels

Neither Zoom nor Google Meet has shipped an equivalent bot-detection policy under that name, but both route third-party assistants through the same chokepoint, because a bot has to be admitted like any other guest. Waiting rooms, authentication requirements, and host-only admission were already the control surface; the AI bot wave has made hosts use them.

Zoom's assistant is unusually explicit about its own presence when sent into a third-party meeting. Zoom documents that "AI Companion will join the meeting as a participant," that the tile carries "Zoom branding and AI Companion indicator, the name of the assistant's owner, activity status (Transcribing)," and that "a chat message is posted 3 minutes after the meeting starts to inform meeting participants of AI Companion's presence and purpose" (Zoom Support: Using AI Companion in third-party meetings). For a candidate that is a notice mechanism. For anyone hoping a bot assistant is discreet, it is the opposite.

Google Meet's Gemini notetaking is first-party and host-initiated, with an on-screen pencil icon and an announcement to participants, and Workspace administrators can gate it by organizational unit. A first-party feature inside the meeting client is not affected by bot-admission controls at all, which is the distinction worth carrying: blocking external bots does nothing to the platform's own transcription.

Three ways an assistant gets audio, and what breaks each one

The Teams change only bites one of these architectures. Comparing them on the same fields makes clear why.

The conclusion holds without the table: an assistant that has to be let into the room can be kept out of the room, and on Teams that is now the default posture. An assistant that never enters the room is not subject to an admission decision, because there is nothing to admit. Control is a desktop application for Windows and macOS that captures the conversation on your own machine rather than joining the call, so a meeting bot policy has no participant to act on. That is an architectural fact about where capture happens, not a claim about what any particular proctoring or assessment system can observe — the criteria for evaluating a desktop AI interview assistant work through the parts that are genuinely verifiable before an interview.

What the policy does not do

It does not stop the interview being recorded. Teams' own transcription and Copilot are first-party features inside the tenant, unaffected by a control aimed at external bots, and an employer that wants a transcript has it without sending anything through the lobby.

It does not resolve notice or consent. A recruiter admitting a bot has been shown a warning about risk; nothing in the policy obliges anyone to tell the candidate what was admitted or how long the output is retained. Those obligations come from law and company policy, not from a meeting setting.

It does not detect everything. Microsoft says so plainly. Absence of a bot label is not evidence that no bot joined.

And it says nothing about assistants that never join. The relevant questions for a local application are different ones — what it captures, what leaves the machine, how long anything is kept — which the AI interview assistant privacy checklist sets out in the order worth asking them.

Before a Teams interview

1. Assume the employer's tenant is on the default, and that a bot-based assistant will sit in the lobby with a label on it. Plan as though it will not be admitted.
2. Read the roster once before you speak. A name you were not introduced to, or a tile with a transcription status, is a bot announcing itself.
3. Ask in writing when you schedule whether the interview will be recorded or transcribed, by what, and for how long. A recruiter who cannot answer has told you something.
4. If you use an assistant on your own side, know which of the three architectures above it is. It determines whether someone else's admin policy decides whether it works.

Which approach fits which interview

If your interview is on the employer's Teams tenant — the common case for a corporate loop in late 2026 — a bot-based assistant is the fragile choice, not because it is prohibited but because its admission is someone else's decision, made while a warning is on screen. A browser-based tool is a better fit for web assessments and a poor one for interviews that run in a locked-down native application. A local desktop application moves the dependency onto your own machine and its permissions, which is a tradeoff rather than a free win: you own the setup, the prompts, and the failure.

Whichever you pick, test it on the platform you will actually be interviewing on, before it matters. Control's free allowance — the first five messages and two minutes of voice for new users — exists for that rehearsal rather than for the interview itself.