This CFO built an automation with AI… and then it broke
This is the third edition of our Bots & Agents series. Last time, I showed you how to give an AI agent the right context. Today, we give it the right workflow.
A lot of finance tasks feel like a never ending story. You're great at your job, but you still finish late because there's so much of it.
So you try to give the whole thing to Claude (or ChatGPT, or Copilot)
Mils (a member of my AI Finance Club) did exactly this. His first Claude skill was built to analyze a sales commissions model. 600 rows and 54 columns, processed every month. The finished skill ran to about 20 pages!
May ran beautifully. Claude even told him he'd done a great job, which he takes with a pinch of salt. But June tripped up on a couple of new GL codes.
To troubleshoot – during our Fractional CFO Corner, Anna Tiomina (one of our AI Finance Club experts) gave him this advce:
Don't build one giant skill, build small ones, then join them up in a project or a plugin.
And that's exactly what she did with her own board reporting. Today it runs on one command.
So today I'll show you how to split a workflow into small skills, one for each type of work, then join them into one plugin.
2 types of work, 2 types of skill
A quick refresher from earlier newsletters.
|
|
AI is probabilistic. Ask it the same question twice and you can get two different answers.
For example, Anna's Claude starts long answers with a TLDR, because she set it up that way. Mils' Claude doesn't.
A formula is deterministic. When you give it the same input, it will produce the same output every time.
Your board report has both kinds of work in it. A KPI has one correct calculation. The variance commentary doesn't.
That's why when you ask AI to carry out one giant task it will struggle.
Whereas small skills let you build each part the right way.
How to build a board reporting plugin in 4 steps
Step 1. Give it the right context first
Before you build anything, set up the folder from our last edition. Four things go in it: a page about your business, your structured data, your unstructured data, and a data guide.
|
|
As an example, for a board report, that means your actuals and your plan as clean tables. Last month's board minutes go in as text files. And your data guide says which file is actuals and which is plan.
Step 2. Map the workflow
Write the full job or task, with one line for every step. Then mark each step: does the answer have to be the same every time, or does it need judgement?
Here's Anna's board reporting
- Pull the data from Microsoft 365 and the client's main system, such as QuickBooks. Same every time
- Check the month is closed, or reconcile the accounts first. Same every time
- Compare the actuals with plan or target. Same every time
- Write the variance commentary. Judgement
- Find useful emails, like meeting minutes. Judgement
- Build the deck in company colors and fonts. Same every time
- Check the deck's numbers against the reporting system. Same every time
Five of the seven steps need the same answer every time. Only two need Claude's judgement.
Step 3. Build one skill for each type of work
Now build each step the right way.
For "same every time" steps, lock the answer.
Tip 1 – Moving lots of data? Use an automation tool:
|
|
Someone asked Anna about whether you can use Claude to move QuickBooks reports into a Dropbox folder.
The short answer is yes – If you can Quickbooks and Dropbox connected to Claude (or another AI) then you can use it to move data.
The question is whether you want to consume your AI credits carrying out deterministic steps like this.
For moving data, it might be better to use an automation tool like Zapier, Make, n8n or Microsoft Power Automate.
This way you can guarantee the steps remain the same every time, without worrying about AI accuracy or your token consumption.
Tip 2 – Ask AI to build the formula once, then reuse it:
|
|
Anna asked Claude to build a skill for a client's rebate reconciliation. Claude said this should be a Python script, because it's a lot of rule-based calculation. So the skill has a script inside it, and the answer is the same every run. She doesn't know Python, but she can still check the outputs.
I say the same thing in every crash course. If you need to calculate something, don’t let AI make up how to do the calculation every time. Work with AI to write the automation, formula, or code once, test it, then reuse it every month.
For judgement steps, tell Claude how to think.
Give it your rules for the analysis, not the answer. Anna's analysis skill tells Claude to start with the top line, then find what drove sales.
Tip 3 – Mostly rules, with a few exceptions? Make Claude ask:
|
|
Anna had a client where one person held all the bank classification rules in her head. They fed Claude old bank statements plus the matching spreadsheets. It mapped roughly 60 to 70% of the rules, and that part became a formula. The other 30 to 40% gets flagged, and Claude must ask before it fills any gap.
That’s what would have saved Mils’ June. A new GL code gets flagged and comes to you, instead of Claude guessing.
Here's how Anna split her board report. Three skills and two agents:
- A main skill that ties it all together. You run it with a forward slash.
- An analysis skill for that company: start with the top line, then find what drove sales.
- A formatting skill for company colors and fonts. Anna calls it one of the easiest to build.
- A deck checker agent that checks the deck's numbers against the reporting system.
- A minutes finder agent that searches emails for anything useful to that month's board.
Step 4. Join your skills into one plugin, then test it
Build the plugin:
1. Go to Customize, then Plugins. Select Add, then "Create with Claude".
|
|
2. Describe your workflow. Paste in your map from Step 2, and say which steps need the same answer every time and which need judgement.
3. Name your tools, like Microsoft 365 or QuickBooks, so Claude adds the connectors.
4. Save the plugin Claude makes. It shows up under Customize, then Plugins.
5. Open it and read the skills and connectors inside. Then sign in to its connectors.
6. Type / in the message box and pick the skill to run your workflow.
Test the plugin
Then test it on a month you already closed. It's November, so run it on September and compare it with what you actually sent.
Automate the work where you know what good looks like first.
Not on Claude? In ChatGPT, a custom GPT holds its own instructions, files and tools. In Microsoft 365 Copilot, you can build an agent in Copilot Studio.
This is where finance is heading. More monthly work will run with an agent in the loop, and our job becomes managing those agents.
But you can only manage one well if you built it from the right pieces.
The One Thing to Remember
One giant skill is hard to build and hard to trust. So split the work by type.
Lock the answers that must stay the same, let Claude handle the probabilistic judgement, then package it all into one plugin.
Then you can work your way up to your 600 row, 54 column sales commission workflow 😉
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 see a CFO assistant built inside Claude? Watch I Added a CFO Assistant Inside Claude Dashboard