This AI Agent Turns Meetings Into Tasks (Claude + monday.com)

What Boards Does The Agent Need In monday.com?

A Projects board and a Tasks board, in what we call a high-level, low-level project management setup. The agent won't work without a project management system like this already in monday.com.

The Projects board holds each project's details (timeline, type, priority, revenue, cost, gross profit) plus four fields the agent relies on: the project name, the client, the point of contact and their email. Those four are how it works out which project a meeting was about.

The Tasks board holds every task for every project. A Connect Boards column links each task to its project, which is how you filter tasks by project, and filling it in correctly is the most important part of the agent's job.

Where Do You Build A Claude Agent?

In the Claude Console, which you reach through the Claude platform. You need admin permissions, then it's Agents on the left and Create agent in the top right.

Each agent needs a name, a description and a system prompt. Rather than write them by hand, we talk the job through with Claude in a normal chat, where we can dictate, and have it write the instructions for us.

How Do Meeting Transcripts Reach The Agent?

Through a folder in Google Drive. Zoom and Teams can record every client meeting automatically and produce a transcript (Zoom's usually arrives about 20 minutes after the meeting ends), and Zapier then drops it into the folder, with a filter so only client meetings get through. OneDrive and SharePoint work too.

We didn't show the Zapier step, for security reasons and because there were no real transcripts to demo with. What we did set up was the folder structure: a Transcripts folder in Google Drive, with a Completed Transcripts folder inside it.

Once a transcript has been turned into tasks, the agent moves it to Completed Transcripts. It only ever looks for new files, so it doesn't waste time and money rereading old meetings or lose track of which ones it has already done.

What Should The Agent's Instructions Cover?

Which boards to use, what each task needs, how to link it to its project and what happens to the transcript afterwards. We dictated the whole brief to Claude. The better you understand the process, the easier this is, and the agent is only as good as these instructions.

For each task: a name, a status, a timeline that acts as the due date, a task type (Claude can check the board for the options) and a long text description of the work. The owner isn't needed in this version.

The key instruction is the Connect Boards column: every task must link back to its project, found using the project name, client, point of contact and contact email on the Projects board. Then the transcript moves to Completed Transcripts, leaving the Transcripts folder empty for the next one.

How Does The Agent Avoid Duplicates & Unlinked Tasks?

Before creating a task, it checks the tasks already linked to that project, so a rerun or the same action raised in two meetings doesn't create a duplicate. If it can't confidently match the meeting to a project, it creates nothing, leaves the transcript where it is and flags it in its report. That beats tasks with no project link.

Claude added both rules when drafting the instructions, plus a default timeline: if nobody mentions one, the task runs from the meeting date to seven days later. The duplicate check could still hit grey areas, so expect some back and forth here and plenty of testing before the agent goes live.

Because our Claude chat was connected to monday.com through an MCP, it could check the board itself. We'd asked for a default status of To Do, but the board only has New, Active, Stuck and Done, so it used New. It also raised Priority and Person columns we hadn't mentioned, and we kept priority and dropped person.

Which Tools & Permissions Does A Meeting Tasks Agent Need?

Two tools: monday.com to write the tasks and Google Drive to read and move the transcripts, both on Always allow. Here's the full setup, in order:

1. Paste the instructions into the summary box and press Generate. The system prompt comes out much simpler than the instructions, which is what we want: the less the agent has to work out, the better its output.
2. Keep the default model, Opus 5.5, on a low effort level. A stronger model or more effort costs more on every run, so it comes down to how often the agent will run.
3. Press Add MCP server and connect monday.com, then Google Drive (or OneDrive), keeping both logins in the same vault.
4. Set both tools to Always allow. An agent that has to ask permission on every run isn't really autonomous.
5. Skip skills, as this agent doesn't need one. Under multi-agent, add an advisor on a stronger model than Opus 5.5 for the agent to consult if it gets stuck.
6. Workflows are optional: they let the agent turn the transcript into a spec and hand each task to sub-agents of its own to create and review. We kept them on as an experiment, knowing it's likely to cost more.
7. Press Create agent. The instructions and permissions can be changed at any point.

One caveat: read access covers your whole monday.com account and your whole Google Drive, with no way to limit it to one board or folder. On a mission-critical system, go through the permissions line by line and deny anything the agent doesn't need. Our accounts were demos, so the risk was low.

How Should The Agent Be Deployed & Triggered?

On a manual trigger while you test it, then ideally by Zapier calling the agent through the API whenever a new transcript lands. Setting up the API side isn't covered in the video.

A schedule also works, say every hour, with the agent checking the Transcripts folder each time and processing anything new. But every run is a session with its own cost, which is why the Zapier trigger is the cheaper option.

The deployment needs an environment, which can be Claude's cloud or your own servers, and we left its networking unrestricted. We also left the budget off while testing, because a session that hits its budget stops before the agent has finished.

What Does It Cost To Turn A Meeting Into Tasks?

It cost $1.04 in our test, and the run took about a minute. The run report showed one transcript in the folder turned into 10 tasks on the Tasks board, which it checked before moving the file to Completed Transcripts.

To test it, we uploaded a sample transcript (a conversation with a test client about their CRM project) to the Transcripts folder, pressed Run on the deployment and opened View sessions to watch the agent work through it live.

The report showed how it matched the project: the test client attended from the same email address as the point of contact, and the client name and project matched too. Every task had a status of New, a timeline, a task type, a priority, the project link and a description ending with a source line, with no duplicates.

What Did The Agent Flag For A Person To Check?

The priorities it chose, an empty person column, some guessed timelines and a transcript dated in the future. We still went through everything ourselves before accepting the run.

Priorities were its own judgement, because we hadn't told it what makes a task critical or medium, so that's a rule you'd need to write. The future date was only because this was a demo, and the guessed timelines were where it had assumed a due date.

The person column was empty on all 10 tasks, and we were glad it said so. You can add rules to assign people, for example based on who was speaking, but assigning people through an agent gets confusing if it isn't done carefully, so we left it out of this version.

Did The Tasks Come Through Correctly In monday.com?

Yes. The transcript had moved to Completed Transcripts, and all 10 tasks were on the Tasks board, each linked to the project with its priority, due date and task type filled in, without us touching anything.

The task descriptions weren't perfect, and how you set tasks up will depend on how your business manages them.

With Zapier delivering the transcripts, the tasks would be in monday.com about 20 minutes after a meeting, for around a dollar. They could be assigned automatically too, and where tasks are consistently the same, agents could in theory start doing some of them, but that's a separate build.

What Would We Refine Before Going Live?

Shorter task names of a few words, tidier task descriptions and the person column filled in.

Changes are made on the agent itself and saved as a new version. The deployment then needs switching to that version and redeploying before it goes live.