Build 3 AI Agents with me [dummy data + guide provided]

Finance pros give me feedback that my workshops are the most practical on AI.

Next week - with Concourse - I am taking this to the next level and inviting you, not just to learn, but to build with me, live!

When you join us here, you will get some data, and a walkthrough guide, and together we will build:

1. AI Cash Dashboard Agent - That cleans data, and reconciles it against the bank. Then builds a rolling cash forecast with a live dashboard
2. AI Collections Agent - That takes AR aging data, ranks every overdue account by how much it's costing you, and drafts chase emails ready to send
3. AI Reporting Agent - That creates an on-brand board deck, auto-generated every Monday. Then writes updates in your own style, and posts to Slack or Teams.

Basically, in 60-mins, you will walk away having built 3 AI workflows.

So save your seat here, and build something to show off in your next meeting ;)

“Review with the team” is not a task

Last week Zsolt from my AI Finance Club posted a month-end close agent in the community using Microsoft Copilot.

It was part of one of our monthly build challenges

The whole thing runs off one Excel tab. There are three columns: task name, the email address of the person assigned, the due date. With a row per checklist item.

This assignment is super important. Because if your close checklist lives in somebody's head (or in a Word document where step four says "review with the team,") the agent has nothing to work with. It cannot email "the team." It needs a name and an address.

You already know this, because you have been the agent.

And every close, you are the one reading the list, working out who is late, and deciding whether to chase them today or tomorrow.

You just never wrote the rules down, because you are the rules!

Zsolt's 3 Building Blocks

Zsolt uses three building blocks is his month-end coordination agent.

IMPORTANT - Zsolt uses an advanced Copilot setup, but the below framworks will work, regardless of which tool you use. Plus we've included examples of other tools you can use for each step.

• Block 1. A team channel, named Finance Month Close, with the people who run the close in it.
• Block 2. A spreadsheet in that channel, with two tabs. Setup, which is your checklist. Log, which is a timestamp and an action, written by the agent every time it does anything.
• Block 3. A task board in the same channel, which holds the tasks once they are created.

Then an agent on top of the 3 blocks, that runs in 2 phases:

1. The setup run: It reads your Setup tab and creates every task on the board with the right person and the right date. In Zsolt’s demo, 3 tasks across 2 people and 3 due dates, plus 3 matching lines in the log.
2. The checking run: It goes through every task, emails each assignee about where theirs stands, emails a status report to the plan owner, and reads the task descriptions for issues people have flagged. In the demo, 2 reminder emails went out (one flagging an overdue task, one carrying an issue somebody had typed into a description) plus 1 status report. That one runs in the morning.

Here’s how you can build this yourself, right now, in any tool.

Your pre-requisite list

You cannot skip this and you cannot automate your way around it. The agent is only ever as good as the specific knowledge it has been given.

Here is the exact prompt I have been using for years to draft a closing checklist that will help you.

"I am an accountant, I need to do a month's closing, can you draft a checklist, it needs to be in a table format and I want to have the following columns: process, account, task, owner and timing. Draft it like you're the most experienced accountant."

Run that, then correct it against what your team does.

Three rules while you correct it.

1. One human name per line. Not a team or a department. If a line has two names, split it into two lines. If a line has no name, you have just found a gap to be addressed ;)
2. Date instead of day number: "WD+3" means nothing to an agent, and if you are honest it means slightly different things to three people on your team.
3. Dependencies: Note them down - "Intercompany reconciliation" is a task. "Intercompany reconciliation, because group consolidation cannot start without it and 4 other entities are waiting on us" is a reason. Nobody chases harder than a person who knows what they are holding up.

Now you have something to build on.

5 Step Month-End Close Agent Build

By the end of these 5 steps you have something that creates your close tasks, chases the people who own them, and reports to you every morning automatically.

I will give you Zsolt's Copilot Studio build as the main example. Then versions in Claude and ChatGPT.

You can also watch Zsolt's demo here.

Step 1. List + Log

Create one file with two tabs. Tab 1 - Setup, is your checklist from above. Tab 2 - Log, has two columns, timestamp and action, and it stays empty. You never write in the Log. The agent does, every time it does something. My auditor brain will not let me skip this, and neither should yours.

Tip - Read the Log before you read the task board. The log tells you what the agent thought it was doing, which lets you fix it if it goes wrong.

Tools - An online Excel file in SharePoint (Zsolt's build), a Google Sheet, or an Airtable base (or other database). All three work. The file just has to be somewhere the agent can connect to.

Step 2. Create your task board

Preferably, in the same place your team already communicates. Again, leave it empty for now, as the agent will fill it.

Tools - Microsoft Planner inside a Teams channel (Zsolt's build), Asana, Trello, Monday, or Todoist. Anything with a task, an assignee and a due date.

Step 3. Give the agent tools

This is the part that surprised me. Zsolt's agent has no custom instructions at all. He turns on something called ‘Generative Orchestration’ and then just hands it capabilities: read the Setup table, read and write the Log table, create a task, get task details, list tasks, send an email, look up a user.

The agent then works out which capabilities to use.

Tip - Give it the smallest set of tools that does the job. Every extra capability is giving it the opportunity to produce something you didn’t ask for.

Tools - In Copilot Studio you configure connectors one by one. This takes a bit more time to set up, but you end up with more control, and less prompting to get the outputs you want.

In Claude you go to Customize -> then Connectors, and add your sheet, your task tool and your email.

In ChatGPT Work they’re found under ‘Plugins’ in the left menu. Claude and ChatGPT require less set up for connectors, but you need to make sure your prompts are good to control your outputs, and get better results.

Step 4. Test it in the chat

Do not schedule anything yet.

Run the setup prompt once and check the result against the log. Zsolt's prompt is this simple: "Please set up the tasks inside the plan based on the Setup table. Please log all actions in the Log table with all details and timestamps." Then check.

Tip - Test with 3 rows, not 30. And test with your own email in both the owner column and the plan owner column, so the first reminders it ever sends come to you.

Tools - Copilot Studio has a test panel built into the agent. In Claude Cowork or ChatGPT Work, run it as a normal task first and ask it to “output in the chat” before you connect it to anything that drafts or sends e-mails.

Step 5. Schedule

Zsolt's second prompt: "Please create a checking run in which you are checking all tasks, sending out reminder emails to all assigned users about the status of the task and also send a status report email to the owner of the plan. Please check all details of the task and look for any issues reported in the description of the task." Then schedule it to run every morning.

Tip - For the first close, have it draft the reminders rather than send them automatically, and send them yourself. An agent emailing your colleagues in your name is a risk, so you need to be sure the drafts are working well, before they’re sent without you.

Tools - Copilot Studio schedules it natively. In Claude and ChatGPT it’s set within ‘Scheduled’ in the left sidebar in both tools.

The One Thing to Remember

Most of us measure the close in days. Five days, eight days. It is the number we report and the number we try to improve.

Next time try this metric. How many chase messages did it take? Emails, Teams pings, the meetings in the corridor (including the ones you sent twice)

This is where an agent can help.

You can’t automate people, but you can automate the coordination, and the chasing.

So this week, if you only do one thing, add a responsible person, due date, and task description to your close checklist. Then get AI to help you

Best,

Your AI Finance Expert,

- Nicolas
​

P.S. - The best way to learn from me directly is to join my weekly masterclasses. 60 mins + Q&A where I'll get all your AI in Finance questions answered. Join me here.

P.P.S. - Want to learn about the wider capabilities of Microsoft Copilot? Check out my latest YouTube video - How To Use Copilot Skills For Finance Like a PRO (2026)​