AI Agents Are Joining Meetings. Start With a Roll Call

The quarterly planning call gains one unfamiliar attendee. Its name looks like a product label, its camera is off, and its role is unclear.

One person thinks it is taking notes. Another assumes it can inspect the project workspace. The host expects it to update a presentation after the call. Nobody knows whether it will remember the discussion tomorrow.

When a new colleague joins a meeting, introductions establish more than a name. We learn why the person is there, what they know, and what part they will play. An AI agent needs the same orientation, with one addition: the room must understand what the software is allowed to do.

That is becoming a practical need rather than a speculative one. On October 7, 2026, Cisco announced Webex capabilities that will let teams invite collaborative AI agents into spaces, meetings, and calls. Its examples include asking one agent to gather usage data and another to update a presentation. The announcement describes forthcoming capabilities, not a universal change to meetings overnight. Still, it makes the direction unusually concrete: an AI can appear inside the place where a group discusses work, then carry some of that work forward.

Before the first substantive question, run an agent roll call.

Attendance is not a permission model

A participant list can tell you that an agent is present. It cannot tell you why it was invited, whose instructions it follows, or whether it can write to another system.

Those distinctions matter because agents can occupy several roles that sound similar in conversation:

• A recorder captures and organizes what happened.
• An adviser retrieves context or suggests questions without changing anything.
• A specialist performs a bounded task, such as analyzing a dataset.
• An operator takes approved actions in other tools.
• A delegate continues work for a named person after the meeting.

Calling all five an “AI assistant” hides the most useful information. The label describes a category of software, not the relationship between that software and the room.

NIST has identified the technical version of this problem. Its 2026 work on software and AI agent identity and authorization focuses on identification, authorization, auditing, and accountability when agents access data, tools, and applications. An ordinary project meeting does not need to reproduce a security architecture. It does need a human-readable account of the same basic facts.

Use a six-line agent roll call

The host or agent sponsor should introduce the agent in less than a minute. Put the same information in the meeting chat or live record so late arrivals can see it.

Identity: What agent is present, and which organization or workspace instance is it using?

Sponsor: Which person brought it into the meeting and is accountable for its participation?

Job: What narrow outcome is it here to produce?

Context: Which meeting content and connected sources may it read?

Authority: May it only answer, may it prepare drafts, or may it take actions in another system?

Exit: When does its access or memory of this meeting end, and who confirms the work is complete?

A useful introduction sounds like this:

This is the analytics agent from our operations workspace. I invited it and own its use today. It may read this meeting transcript and the approved usage dashboard to compare three launch scenarios. It can answer questions and place a draft table in the project folder, but it cannot message customers, change source data, or update the final deck. Its access to this session ends after the draft is reviewed tomorrow.

That statement is not legal theater. It lets every participant adjust intelligently. Someone can object that the dashboard contains restricted customer data. The analyst can correct the source. The host can explain that the call includes an external partner, so the agent should use an aggregated view. Ambiguity becomes visible before it becomes action.

Name one human sponsor, even when many people can prompt

If everyone can ask the agent for work, it may feel collectively owned. Accountability then becomes strangely private when something goes wrong.

The sponsor is not required to type every prompt or approve every harmless response. The sponsor makes three decisions on behalf of the room: whether the agent belongs in this meeting, whether its role has changed, and whether its output is ready to leave the room.

This prevents a subtle form of scope drift. A research agent joins to compare pricing options. Midway through the call, someone asks it to draft customer language. Ten minutes later, another participant asks it to place that language in the email campaign. Each step sounds adjacent to the last, but the agent has moved from analysis to external communication.

The sponsor should interrupt that chain with a plain sentence: “That is outside today’s role. We can assign it separately with the right owner and approval.”

Reintroduce the agent when its role changes

People renegotiate meeting roles naturally. A finance partner who joined as an observer may be asked to validate a forecast. An engineer may become the decision owner when the conversation turns to reliability. The room notices because the person speaks and the host acknowledges the shift.

Agent role changes can be quieter. A new connector is enabled, another participant grants access, or a read-only task becomes a request to update a file. Treat any change in context, authority, audience, or retention as a reason for a second roll call.

Keep it short:

We are expanding the analytics agent’s job to draft the scenario table in the project folder. It still cannot alter the source dashboard or publish the presentation. I remain the sponsor.

This is especially important when the audience changes. If a customer, candidate, vendor, or employee joins midway, the agent’s original context boundary may no longer fit the room. Pause before letting connected information flow into the new conversation.

Give the agent’s output a visible status

An agent can produce something during a meeting that looks finished long before it has been checked. A well-formatted table, summary, or recommendation acquires authority from its appearance.

Use one of three statuses:

1. Working: The agent is still gathering or transforming information. Do not rely on the output.
2. Ready for review: The requested artifact exists, but the named reviewer has not accepted it.
3. Accepted: A person with the right expertise and authority has checked it for the intended use.

The status should travel with the artifact, not remain in someone’s memory. “Ready for review by Mina” is more honest than a clean file named final-scenarios.

Review is not a ceremonial human click. The reviewer should know which sources the agent used, which important sources it could not access, and whether it took any action beyond producing the visible output. NIST’s August 2026 workshop summary describes related work on binding accountable people to agent actions so organizations can audit and understand chains of trust. The meeting habit is simpler, but the principle travels well: make the path from instruction to consequence inspectable.

End with checkout, not disappearance

Human attendees often leave quietly. An agent should check out explicitly because its work may continue after its icon vanishes.

At the close, answer four questions:

• What did the agent produce or change?
• What remains in working or review status?
• Where can participants inspect the output and source record?
• What access, scheduled work, or retained context continues after the meeting?

If the agent will keep working, name the sponsor and the next checkpoint. If it will not, end the session access and record that fact. “The agent left” is not enough when a background task, reusable token, or connected workflow remains active.

Make machine participation legible

The value of an agent in a meeting is not that it can imitate another face in the grid. It can bring relevant context closer, perform a bounded piece of work, and reduce the distance between discussion and follow-through.

But useful participation depends on legibility. People should not have to infer an agent’s sponsor from who first mentioned it, its permissions from what it happens to attempt, or its memory policy from a settings page they have never seen.

An agent roll call gives the room a shared model before the work begins. Name the identity, sponsor, job, context, authority, and exit. Reintroduce the agent when any of those change. Then close with a visible account of what it did.

Caspi supports meetings with live recap, suggested questions, contextual chat, proactive flags from connected tools, post-call action items, and persistent meeting memory. As meetings include more connected AI participants, that real-time record can help a team keep each agent’s role visible alongside the human conversation.