AI Training for CFOs: What to Learn, in What Order
A CFO needs five pieces of AI training, in this order: the rules on what data goes into a model, daily use of one model on your own numbers, writing a rulebook the model loads every time, building one scheduled agent with read access to the ledger, and the controls and cost caps that sit around it. Do it on your own close and your own board pack, not on sample data, and it takes about 12 weeks at 3 to 4 hours a week to have one agent running.
Why most CFO AI training changes nothing in the close
Most CFOs I coach have already done a course. I ask one question: what changed in your month end afterwards? The usual answer is nothing. The course ran on a sample company, the exercises were prompts about imaginary invoices, and on Monday they went back to a 9 day close with the same spreadsheets.
The second mistake is the order. People start with the clever part, building agents, before they've decided what data is allowed into a model. Then someone on the team pastes the payroll file into a free chat account, the audit committee hears about it, and AI gets banned in finance for a year. Learn the rules first and the rest is safe to try.
The 12 week order I use with a CFO
| Stage | Weeks | What you learn | Proof you have it |
|---|---|---|---|
| 1. Data rules | 1 | What may go into which model, on which plan | A one page policy your team has read |
| 2. Daily use | 1 to 4 | One model on your own numbers, every day | 20 real finance tasks done with it |
| 3. The rulebook | 3 to 8 | Writing rules the model loads, and checking its answers | Variance commentary you'd send with light edits |
| 4. One agent | 9 to 12 | A scheduled run with read access to the ledger | 4 weeks in shadow at 95% agreement |
| 5. Controls and cost | 12 on | Sign-off points, logs, run and token caps | An auditor could follow every action |
1. The data rules, before any tool
Spend the first week on one page, not on prompts. It says which data can go into a model and on which plan: public filings anywhere, your own management accounts only on a business plan where the provider doesn't train on your inputs, payroll and personal data nowhere until you've had legal sign it off. I write this with the CFO in about 2 hours. It's the page that lets the team say yes to everything else.
2. Daily use of one model on your own numbers
Pick one model, Claude or ChatGPT, and use it every working day for 4 weeks. Use it on real work: last month's variance analysis, a first draft of the board pack narrative, a lender covenant you need to reread. I set a target of 20 real tasks. By task 20 you know where it saves an hour and where it makes up a number. You can't learn that from a course.
3. Write the rulebook, then check every answer
This is the skill that separates a CFO who uses AI from one who runs it. You write down what the model needs every time: the chart of accounts, your materiality threshold, how you describe a variance to this board, the five words the chair hates. It goes in a project or a rules file the model loads on every run. Then you check its output the way you'd check a junior's reconciliation, line by line against the ledger. Expect the first drafts to need 30 to 40 minutes of edits. By week 8 it should be under 10.
4. Build one agent, read only
Now you build. One agent, one task, on a schedule. My usual first one matches the bank feed to open invoices overnight and leaves a list of what it couldn't match. It reads the ledger and the bank, it writes nothing back. Run it in shadow for 4 weeks next to the person who does the job today. I want 95% agreement over 2 straight weeks before it posts anything. You don't need to write code for this. You need to read a log and edit a text file.
5. Controls and cost
The last piece is the one your auditor will ask about. Which actions does a person sign off? Where is the log of what the agent did and why? What's the weekly cap on runs and tokens, and who can raise it? My rule is that an agent prepares a payment and a named person releases it, and that line doesn't move after 6 good months. Once you can answer all of that, you're ready to teach your team, which takes another quarter.
From my operating seat
I'm Hayat Amin, an AI-native CFO, AI operator and AI coach for CFOs who teaches finance leaders how to build AI into finance operations. I spent twenty years in the C-suite, with three exits and three FT100 listings. Now I build and run AI agent systems inside companies, across finance, sales pipeline, onboarding and compliance evidence, and I coach CFOs one to one on the same work.
I teach it in this order because each stage needs the one before it. Inside one client I run, an agent posts a live cash runway figure before the finance team logs in each morning. I don't let anyone rely on that figure until they've checked it against the bank every morning for 4 weeks. That's the training: your numbers, your rules, and a check you'd sign.
What is the best AI course for a CFO?
The one that runs on your own numbers and ends with something live. Judge any course or coach on three tests: you work on your own ledger and board pack from week one, you finish with a scheduled agent doing a real finance task, and someone checks its output against the books with you. A course that teaches tools on sample data passes none of them.
Do CFOs need to learn to code to use AI?
No. A CFO needs to read a run log, edit a plain text rules file and check a model's answer against the ledger. That's the same skill set as reviewing a reconciliation. Code matters once an agent sits inside your own product or runs at volume, and by then you hire one engineer part time to own it.
How long does it take a CFO to learn AI?
About 12 weeks to one agent running on a real finance task, at 3 to 4 hours a week. The first 2 weeks are the data rules and daily use. Weeks 3 to 8 are the rulebook and checking output. Weeks 9 to 12 are the first agent in shadow. Teaching your team takes another quarter.
Where I come in
This is the work I do with CFOs. We start with your data rules, use your last close as the training set, and finish with one agent running on your ledger that you've checked yourself. The full map is at AI for CFOs, what the end state looks like is at the AI-native CFO, and the one to one sessions are at coaching for CFOs. Which part of your close would you want an agent on first? Start at meethayat.com and tell me.