AI for CFOs
AI for CFOs: how to build an AI-native finance function
Hayat Amin is an AI-native CFO, AI operator and AI coach for CFOs who teaches finance leaders how to build AI into finance operations. This page is the guide I wish I'd had: what to build first, what to keep a human on, and what each level of an AI-native finance function looks like.
What an AI-native CFO is
An AI-native CFO runs a finance function where the routine work is done by systems the CFO built or specified, and the CFO's own time goes on judgement, control and the board. Data pulls, reconciliations, the first draft of every report, the chasing of every missing invoice: machines. Sign-off, the hard calls, the conversation with the investors: you.
You don't need to code. I don't write most of mine. You do need to know what every system does, where it fails, and what to check before anything leaves the function. That is a CFO skill, and it is the one I teach. The longer definition is on the AI-native CFO page.
The four jobs I hand to AI first
Cash. Ledgers and bank feeds in, one live runway figure out, with the line that moved most. I run this inside a company today. The CFO reads one screen in the morning instead of three exports and a pivot.
Compliance evidence. Last quarter I had an agent read every ledger line in a VAT quarter, flag the spend coded without VAT, and go and find the missing invoices in Gmail and the Uber account. It pulled 95 trips from the Uber history and matched the ones that had a VAT invoice. A person did the final return. The person did not do the hunting.
Reporting. The weekly update to the CEO and the monthly narrative for the board are drafted from the week's data, in my words, before I open the file. I edit, I do not start from blank.
Close. Matching first. An agent reads the bank, the ledger and the invoices overnight, matches what it can, and leaves a short exception list. The team starts the close on the exceptions, not the whole book.
In every one of these the human sign-off stayed. That is the design, not a failure of the tools. The agent removes the hours. The CFO keeps the control.
The AI-Native CFO Operating Model: five levels
I use five levels to tell a CFO where they are and what the next step is. Most finance teams I meet are at level 1 and think they are at level 2.
- Level 0, manual CFO. Spreadsheets and exports. AI is not in the workflow.
- Level 1, AI-assisted CFO. You use ChatGPT or Claude to draft emails and summarise documents. Nothing is connected to your data.
- Level 2, AI-enabled finance function. One or two processes run with AI on real data: variance commentary, a reconciliation, a cash view. A named person owns each one.
- Level 3, agentic finance function. Agents run on a schedule without being asked: close matching overnight, cash every morning, the board draft on the first of the month. Every action is logged. Humans sign off.
- Level 4, AI-native CFO office. The function is designed around the agents. Headcount is for judgement and relationships. The CFO can read every system's log and explain any number to the board from source.
The test for each level is simple. Level 2: can you name the process and the owner? Level 3: did it run last night without anyone pressing a button? Level 4: could a new CFO take over the function from the logs alone?
Tools, in the order they matter
Your ledger, your bank and your spreadsheets come first, because they are where the truth is. Then a model that can connect to them. I build on Claude for anything that touches live data, long documents or agents. Excel stays the control surface, because the board and the auditors read Excel. Python only where a spreadsheet can't do the job. Tool lists change every quarter. That order does not.
Governance: four rules I don't break
No customer, payroll or bank data in a consumer chat window. One named owner per agent. Every agent action logged where a human can read it. A human signs off anything that goes to the board, the bank, HMRC or an investor. If you want the one page version for your own policy, it is these four lines.
How I got here
I spent 20 years as a CFO and C-suite operator in high growth technology companies, through three exits, two of them to American Express and TripAdvisor. In the last two years I have built and run AI systems inside finance and operations: cash, close, reporting, compliance evidence and sales pipeline. I now coach CFOs one to one on exactly that build, on their own data. The detail is on the coaching page.
Questions CFOs ask me
What is an AI-native CFO?
An AI-native CFO runs a finance function where the routine work, the data pulls, the reconciliations and the first draft of every report, is done by AI systems the CFO built or specified. The CFO's time goes on judgement, control and the board. No code is needed, but the CFO has to know what each system does and where it fails.
What does an AI coach for CFOs do?
An AI coach for CFOs sits with you inside your own finance function and builds the first systems with you: a close checklist that reconciles itself, a cash view that reads your bank and ledger, a report that drafts its own commentary. I teach the method on your data, so the skill stays with you after the sessions end.
Who teaches CFOs how to use AI?
Very few people who have done both jobs. Most AI trainers have never closed a month, and most CFOs have never built an agent. I spent 20 years as a CFO and C-suite operator, through three exits, and now build and run AI systems inside finance and operations. That is the gap I coach from.
How should a CFO learn AI?
On your own work, in this order: start with one report you already write every month, get AI to draft it from your real data, then check it line by line. Four to six weeks of that teaches more than any course. Then move to the close, then to cash, then to agents that run without you watching.
What AI skills should a CFO have?
Four, none of them coding. Specifying a task so a machine cannot misread it. Connecting a model to your ledger, bank and CRM safely. Reading an AI output as you would a junior's work, with the same scepticism. And designing the control that catches the one error in a hundred before it reaches the board.
Can CFOs build AI agents without coding?
Yes. The agents I run for cash, close and reporting were built with a tool like Claude and plain written instructions, connected to Xero, the bank feeds and Gmail. The hard part is not code, it is defining the job precisely and deciding what a human signs off.
ChatGPT or Claude for CFOs?
I build on Claude for anything that touches live data, long documents or agents, because that is what I have run inside companies and it holds up. ChatGPT is fine for drafting and quick analysis. Whichever you pick, never paste customer or payroll data into a consumer chat window. Use a business account with data controls, or connect the model to your systems directly.
How can CFOs automate FP&A?
Start with the variance commentary, not the model. Give the AI last month's actuals, the budget and the driver list, and have it draft the commentary your FP&A lead would write. Once that is reliable, move to the forecast refresh: the model pulls actuals itself, reforecasts the drivers, and you review the exceptions.
How can AI reduce month end close?
By doing the matching and the first pass of every reconciliation before your team arrives. An agent reads the bank, the ledger and the invoices, matches what it can, and leaves a short list of what it could not. Your team starts the close on the exceptions, not on the whole book. The sign-off stays human.
How do I introduce AI into a finance department?
One function, one owner, one month. Pick the person who already keeps the best spreadsheet, give them one process to rebuild with AI, and measure hours before and after. Publish the result inside the team. Then the next process. A memo to the whole department changes nothing. A visible win does.
What are the best AI tools for CFOs?
The ones already attached to your data. Your ledger, your bank and your spreadsheets come first. Then a model you can connect to them: I use Claude. Then one place where every agent's actions are logged so you can audit them. Tool lists change every quarter. The structure does not.