Build Your First AI Finance Agent in Claude (No Coding)

0:00AI agents are becoming one of the biggest opportunities in finance, but for most finance professionals actually building one still sounds far more technical than it needs to be. Because the real opportunity isn't having another AI you constantly have to prompt. That's not the point. It's The real point is being able to hand off parts of your own finance work and let an agent work through them for you. Or maybe you want to give it an entire finance responsibility the same way you would give it to someone on your team and then have it come back to you with what needs your attention. So in this video, I'm going to show you how to

0:36build your first AI finance agent in Claude from scratch without writing a single line of code. And we'll go from defining what you want your agent to do all the way to having it run autonomously and then seeing how it handles a real finance workflow on its own. Now, one quick thing before we start. I'm about to launch a complete program designed to help finance professionals become AI native and transform how they work with AI. And the waitlist is now open. Everyone who joins will get access to the lowest price we will ever offer and five people will

1:12receive an annual membership for free. To join the waitlist through the top link in the description below. But for now, let's get into it. So what is an AI finance agent? Well, an AI finance agent is an AI system with a defined finance responsibility, the business context and also the tools to handle it and then enough autonomy to decide how the work gets done. In normal Claude conversations, I might upload a P&L or ask for the biggest variances and then decide what to investigate next. I'm directing every single step, but with an agent, I define the job once, monitor our financial performance, find what needs attention, investigate it, and

1:58then report back. And then the agent decides where to look and how far to dig based on what it finds. That's also what makes it different from a scheduled prompt. A scheduled prompt follows the same steps every time. An agent adapts its investigation to the situation, but it still works within the boundaries that I define, like what counts as material, what evidence it needs, and when something needs human review. So, the simplest way to think about it is a digital finance worker with a clear job, with the knowledge to do it, and also

2:37with the rules that it has to follow. And the first thing we need to do is to define that job. Choose your agent's job. The job has to be a clear business objective. For this video, I'm going to build a financial monitoring agent for an example company Evergreen Industrial Group. Now, its job is to review the latest completed month, identify any material issues or unexpected developments, investigate the ones that deserve attention, and then report what it finds. I keep that job broad, and broad on purpose. I do not tell it to check gross margin or to look into manufacturing costs. Say the month shows revenue slightly below budget, margin

3:24down in one product family, higher manufacturing costs, and a KPI that moved sharply. The agent has to then decide which of those is material, and then where to look next. And that decision belongs to the agent. The process it follows is review, identify, decide, investigate, conclude, and then report. Now, the job ends with an output. Every run produces a finance exception report that covers what happened, why it matters, the likely driver with its financial impact, and also, critically, what needs human review. If nothing material happened that month, it still produces the report and says so. It never manufactures a problem to make the output look interesting. So, with the job defined,

4:16the agent needs the data to actually do the work. Give your agent the financial data. An agent can only investigate what it can reach. So, I start with the data. So, I have a folder in Google Drive called Evergreen Finance, and it holds the kind of monthly reporting package our finance team works from. The monthly P&L is where every review starts, and it runs from revenue and COGS down to EBITDA with the current month against the prior month and budget. Behind it are five supporting files. Budget versus actual shows performance against plan by product family. Revenue by product breaks sales into volume, selling price, and customer. Cost detail goes into

4:57materials, labor, overhead, and freight. The KPI file holds the operational measures, and the general ledger file holds the account balances and journal entries underneath everything. The agent starts at the P&L and moves into these files only when something deserves a closer look. Build your AI finance agent. Now, the agent needs somewhere to live, and in Claude, that is a project. Now, a project is a workspace that keeps files and instructions available to every task that run inside it. That way, the agent does not have to relearn the company on every single run. So, I click projects, which opens the project workspace. And from there, I click new

5:41project. I name it Evergreen Financial Monitoring Agent and add a short description of what the agent does. Then, I give the project one standing instruction. The context files are the authoritative source for everything about Evergreen, and Claude must never invent or override what they say. Now, I drag six files into the project context at the same time. Five are context files, and the sixth is agent.md, which I will come back to once the other five are clear. The first two explain the business and the numbers. Company context covers what Evergreen makes, its product families, its customers, and its cost structure. Finance context covers how the data should be read, including

6:29the variance calculations, the KPI definitions, and how the supporting files reconcile back to the P&L. The third file shapes the agent's and behavior more than any other, so it's worth highlighting this. It is the materiality and investigation framework, and it tells the agent how to decide specifically what deserves attention. So, it weighs the size of a movement, its business significance, whether it

6:56persists, and how reliable the data is. And it also sets how deep an investigation goes, and how to separate a fact from a hypothesis. And without this file, the agent basically has no basis for choosing. I would get one long variance report where every line looks equally important. The last two files control the output. The finance exception report specification defines what the report has to contain, and it sets the sections every report needs, from the executive summary and the material findings to the financial

7:34impact and the items for human review. The email and report formatting framework defines how that analysis is going to be presented and it covers the short executive email, the detailed PDF attached to it, and the formatting rules both follow such as terminology, number conventions, and source references. Context tells the agent about the company. It also needs a method for the analysis itself, and that comes from a skill. Now, mine is a financial analysis skill, and it has the agent validate the data, connect material movements to the operational drivers, check alternative explanations, and then verify the calculations before it concludes anything. Now, I built it with my craft

8:24framework, which is the process I use to turn finance expertise into structured AI skills that can perform the work reliably. Now, that leaves the sixth file, which is agent.md. Now, the context files are knowledge, and the skill is capability. This file is the operating instruction. It tells Claude what it is responsible for, which resources it can use, and how it should work. It opens with the role. Claude is the financial monitoring agent for Evergreen Industrial Group. It reviews the latest completed month, decides what deserves attention, investigates it with the available evidence, and then

9:07produces the finance exception report. Then, it connects the pieces. The five context files are the authority on the company, and the financial analysis skill handles the analysis. To put the data and delivery, it names two connectors, which are links between Claude and an outside app. The Google Drive connector gives the agent the live

9:31files in the Evergreen Finance folder. The agent is told to review whatever files are currently in that folder instead of a fixed list of file names. Then, when next month's files replace this month's, the same agent keeps working and I don't have to rewrite anything. Zapier is the action layer. So, once the report is finished, the agent uses it to send the email through Gmail. Before that email goes out, the figures in the email and the attached report have to actually match. The behavior section is the part to read closely. The agent starts with the P&L and identifies the important movements, and then it decides for itself which

10:12ones to investigate. So, it uses the supporting files to dig in, cross-checks its conclusions, and then quantifies the financial impact. It is also told directly never to begin with a predetermined problem or to assume that revenue, margin, or costs are the issue. The file carries the limits as well. Conclusions need evidence. Facts stay separate from hypotheses, and the agent stops where the evidence stops. Anything it cannot resolve goes into the report as an item for human review. With the six files and the skill in the project, the agent is now built. Now, if you want agent.md, the context files, and the demo data from this video, they're all

10:57in my free community, and the link is in the description below. Right now, the agent only runs when I start it myself, and I don't want to have to come back to every month do that, and that is where a scheduled task comes in. So, now let's go to automating the agent. Automate your agent. From the project workspace, I go to the scheduled section, and I click the plus button. I name the task Evergreen Financial Monitoring. The instruction is short. [snorts] It points Claude at agent.md and tells it to run the complete monitoring workflow through to the email. Now, it can stay short because the logic already lives in

11:37agent.md and in the project. I set the frequency to monthly on the 5th at 9:00 a.m. and the agent has to run after the month-end data has landed in the folder, so the date is chosen around the close. Under permissions, I select skip all approvals. If approvals are required, then the run pauses at the first action that needs permission and then it just waits. Nobody is there to approve it at 9:00 in the morning on the 5th, so then the report just doesn't go out. Skipping approvals lets the workflow run to the end. The analytical rules stay exactly as they were because those come from the

12:14context files, the skill, and agent.md. I click save and the task shows as active, then it's linked to the project and then it shows the monthly recurrence and the next run date. The schedule is only the trigger. It decides when the agent wakes up. What happens after that depends on what the agent finds in that month's numbers. It even reads the reporting period from the files themselves, so nothing in the task changes from month to month. But, I do not want to wait until the 5th to find out if it works, so I test it now. Test your agent. To do so, the task has a run

12:53now option. So, I click it and the full workflow starts exactly as it would on the scheduled date. So, the run shows 29 commands, two integrations, and seven tools. And before it touches any data, the agent reads agent.md and the five context files, and then it loads the financial analysis skill. It then learns the company and the rules first. Then it goes to Google Drive. It searches the Evergreen finance folder, reads the files, and downloads the workbooks so it

13:22can work through the underlying sheets. And from there, I can watch it move. It goes from the P&L into cost detail, revenue and the KPIs, then into the general ledger and the August journal entries. And once the analysis is done, the agent then builds the report. It renders a PDF and inspects the pages visually before it then moves on. And then it uses Zapier Gmail action to send the email with the PDF attached. Now, I open the email first. The subject line carries the company, the report type, the period, and the status, which is human review required. The body here opens with a short executive summary and

14:02a key figures table that shows revenue, gross profit, and EBITDA against the prior month and budget. And below that are the material findings with their amounts, followed by the items that need human review. So, a finance leader now knows what needs attention before actually even opening the attachment. Then, I open the attached PDF. It is an 11-page finance exception report for August 2026. The first page gives the same status and executive summary, followed by the overall [snorts] financial picture against July and budget. And after that, it goes through each finding with its calculations, the journal evidence, and the alternative explanations the agent considered. Now, the first finding is an alloy steel

14:48sourcing issue. Evergreen bought 1,600 tons of alloy steel from a non-contract supplier at a price well above standard with expedited freight on top. The agent traces the purchase price variance and the freight through to an $817,000 hit to gross profit against budget. That one issue explains 99.2% of the gross profit variance. The second is a credit loss allowance for a

15:20customer called Halverson Machinery. Credit loss expense came in at $740,000 against a budget of 90,000. Most of it a specific allowance booked through a manual journal entry. And the agent then connects that to two other signals it found. Receivables more than 90 days past due went up and Halverson's revenue went down. So, the report is careful about what it claims. On the one hand, it confirms the amount and the nature of the charge. And it also states that the files are not enough to judge whether the allowance is adequate. So, the question now goes to a human, which is what we want. It does the same on the

16:03steel issue where it cannot say why contracted supply fell because supply contracts were not in the folder. Then there is something small. It finds two positions of $22,400 to Cloudline Software, same vendor and same amount two days apart. The amount is immaterial. The pattern looks like a possible duplicate invoice, which the framework treats as a trigger. So, the agent then flags it for accounts payable instead of just ignoring it. And the other smaller movements are reviewed, classified, and then left there so the report stays on what really matters. The report closes with three human review items each with an owner and a question to answer. The steel issue goes to

16:51procurement, the Halverson allowance goes to the controller, and the possible duplicate to accounts payable. Now, to check a run like this, I go to the validation checks recorded in the report. Product family revenue ties to total revenue. COGS detail ties to the P&L and the general ledger net result agrees to net income. Then I compare the email against the PDF because the period, the figures, and the review items have to match. And that is the full build. A project holds the company context, and a skill handles the analysis. Two connectors cover the data and the delivery. One agent file defines the job, and a schedule wakes it up each

17:37month. Your role changes once this is running. You are no longer the one pulling files and chasing variances. You open an email, you read what the agent found, and then you decide what to do about the items it could not decide for you. Now, if you want to go further and learn how to become AI native and transform how you work with AI across finance, that's exactly what the program is designed to help you do. The waitlist is now open through the link in the description, where you'll also get access to the lowest price we will ever

18:08offer. And I'll see you in the next one.