JPMorgan saved 360,000 hours with one AI tool – Here’s the method behind it:

I want to tell you about a session we ran recently in the AI Finance Club.

We gave 60 finance professionals a dataset for a fictional company and set them a challenge.

Build an AI-powered solution to reduce month-end – In one week – with whatever tool you want.

What happened next was not what you'd expect.

Within 48 hours, teams of finance managers, controllers, and fractional CFOs (most had never built anything with AI before) had working prototypes.

• One team built a tool that flagged journal entries missing approval status before the close even started.
• Another automated their GL exception review: a process that previously took 45 minutes every morning was running in 90 seconds.
• A third cut their monthly forecast update from 4 hours to under 20 minutes.

The method that made that possible isn't new.

It's the same method JPMorgan used to save 360k hours. It's the same framework Deloitte uses in its digital transformation engagements. And the same one Microsoft's team used to ship GitHub Copilot (which now has 4.7 million paid subscribers and 20 million total users).

And it's what I want to share with you today.

​

​
The Waiting Game

I speak with hundreds of finance professionals every month, and I see the same thing.

You know AI could help you with something in your function. And, you’ve thought about what that something is. But you haven't built anything yet. And this is costing you.

While you're working out where to start, other finance teams are building.

And, the gap between finance functions that build and those that don't is getting bigger every day.

Often the reason is technical: "I don't know how to code." Sometimes it's red tape: "IT would need to approve it." Sometimes you just don’t know where to start.

But, the problem is actually that you don’t have a clear method.

Using methods like the one I’m going to show you helped JPMorgan build a tool called COiN (Contract Intelligence).

It reviews 12,000 commercial credit agreements in seconds. The same work previously took lawyers and loan officers 360,000 hours a year to complete manually.

They found ONE problem, and solved it.

The same principle applies to your reconciliation. Your GL review. Your management reports.

Without needing coding skills, without needing IT sign off (you already have the tools) and without taking months of planning. You can use this method now.

It's been used in software development for years. And it works even better for finance professionals than for engineers, because finance pros already understand the process they're trying to improve!

​

​
Building Something Imperfect

The idea comes from what's called the MVP that stands for Minimum Viable Product.

Instead of building the perfect solution, you build the simplest possible version that solves the problem. Then you test it, improve it, and expand from there.

It's the operating model behind some of the most successful products ever shipped.

• Microsoft launched GitHub Copilot with a single feature: autocomplete for code. It now has 4.7 million paid subscribers and 20 million total users.
• OpenAI launched ChatGPT with no additional. Just a chat box. And, within two months it was one of the fastest-growing consumer applications in history.
• Anthropic shipped Claude Artifacts (the feature that lets you build interactive tools directly inside a chat window), as a simple prototype before expanding it into one of the platform's most-used features.
• Deloitte runs rapid prototyping sprints (part of its "Greenhouse" methodology) where complex business problems are turned into working prototypes in days instead of months.

The finance professionals in our AI Finance Club hackathon who produced the best results did exactly the same thing. They picked the smallest problem they could define, and built something that solved just that.

If you've never finished building an AI tool, it's almost always because you tried to build everything at once.

You wanted it to handle all 12 reconciliation tabs, all 5 legal entities, every use case. So you never finished – because you can't do all of that in 5 days.

But you can do ONE of them.

​

​
How to Do This in 5 Days

Day 1: Write the problem down in precise terms

Before you open any AI tool, write one paragraph that answers these 4 questions:

• What is the specific finance task I want to improve?
• What data goes into it?
• What does the output look like?
• Who uses that output and for what decision?

The more specific, the better. "I want to improve month-end close" is not specific. "I want to automatically flag journal entries in our GL that are missing an approval status, before the formal close starts" is specific.

One of these can become a working tool in 5 days. The other cannot.

This is the same principle that makes AI outputs better in general: more context, better result.

You are writing the brief for your own build.

Day 2: Design the solution

On Day 2, don't build anything yet. Instead, sketch what the tool needs to do, step by step. What does it receive as input? What does it do with that input? What does it produce?

Then take that sketch and paste it into Claude or ChatGPT with this prompt:

Use the simplest option that would still be genuinely useful. You're not designing the final product. You're designing the prototype.

Tool Spotlight: Claude Projects In Claude, you can create a Project (a workspace where your context is always loaded). Add your company's financial structure, your reporting cadence, your ERP setup, and the specific problem you're solving as the Project instructions. Once you do this, every conversation inside that Project starts with your full context already active, so you never have to re-explain your business. Use it on Day 2 to design your solution, and keep using it through to Day 5 to build, test, and iterate. How to set it up (takes 10 minutes): 1. Open Claude → click "Projects" in the left sidebar → "New Project" 2. Add a project name (e.g. "Month-End Automation Build") 3. Click "Add project instructions" and paste in your company context: entity structure, ERP, the process you're automating, and any constraints (e.g. "we use Excel, not Python") 4. Every conversation you start inside this project will have that context from the first message This works particularly well for finance because your processes are consistent and repeatable. Exactly the kind of context an AI benefits from having loaded permanently. It's one of those features that, once you start using it, completely changes how you work with AI. ​ ​Pro tip – When you end a conversation in a project, ask: Create an artifact giving a detailed summary of the conversation we’ve just had. After that, you can add the artifact to the project knowledge. So every time you revisit, the context gets better and better.

Day 3: Build version 1

Open your preferred AI tool and start building. Options for finance prototypes are:

• A Claude artifact: Good for HTML dashboards and interactive tools that run in the browser without any setup.
• A custom GPT or Gemini Gem – Good for conversational interfaces that guide someone through a process using a chat format.
• A Python script in Google Colab – Good for processing large datasets, running forecasting models, or producing visuals you cannot in Excel (e.g Heatmaps).
• An Excel macro (VBA or Office Script) – Good for automating repetitive tasks within your existing Excel or Office workflows.

Choose based on what you and your team will use, not what sounds most impressive. A working Excel macro your team runs every month is worth more than an impressive Python script that nobody uses.

Day 4: Test on real data and fix the 3 biggest problems

Run your prototype against real (or realistic dummy) data. Not the clean, well-formatted example data you used while building.

Use the messy, real-world data your team works with. So that you can note what breaks, what's confusing, and what surprises you.

Fix the 3 biggest issues. You don’t need to fix all of them. Just the 3 that most prevent the tool from being useful.

This is the hardest thing with the MVP method- Not fixing every single thing before you start using it.

Day 5: Demo it and document it

Show the working prototype to one person outside the team who experiences the problem it's solving. Walk them through what the problem was, what the tool does, what the output looks like, and what still needs improvement.

Record a short video of the demo. This serves two purposes.

First, it becomes your documentation: Anyone on the team can understand how the tool works by watching the video.

Second, it's your internal pitch if you want to scale this to the other areas of the business.

Then quantify the before and after as specifically as you can:

"This task used to take 45 minutes manually. The prototype does it in 90 seconds. There are still 3 edge cases it can't handle, and I've documented those."

It might not be perfect, but if you can free up a little more time, and make a faster decision.

The next time you build you will free up more time, and make even faster decisions.

The more you repeat, and iterate over time, the more you improve, the more impressive tools you will build.

The One Thing to Remember

The finance professionals in that AI Finance Club hackathon who produced the best results were not the ones with the most technical background.

They were the ones who picked the smallest, most specific problem they could think of and built something that solved that one problem completely.

You don't need a data science team. You don't need IT approval to build a prototype.

You need a clear problem, your existing tools, and 5 days.

What will you build?

​

Best,

Your AI Finance Expert,

Nicolas

​

P.S. – What's the one finance task you've been thinking about automating? I want to know what's on your list. Hit reply (I read all replies)

P.P.S. – If you're new to creating custom AI assistants, I have covered this in one of my recent videos about ChatGPT's Custom GPT feature. You can check out my video here: My Secret to Create CustomGPTs for Finance in 2 Minutes.

​

​link​