Guide
AI for CFOs
Every CFO has now been asked some version of the same question: what are we doing with AI in finance? The honest answer depends on a distinction most AI tools blur. A finance team does two kinds of work. One is producing numbers. The other is standing behind them. AI is getting good at the first. The second is still the CFO's job, and it's where most AI output fails.
CFO Sentinel is built for the second kind of work. It is a financial-intelligence desk that turns a question and a file into a board-ready record. Facts are separated from assumptions, cash is separated from EBITDA, and the jurisdiction is named. Human judgment remains the authority.
What CFOs actually need from AI
Speed matters less than people think. A first draft of variance commentary in thirty seconds saves an analyst an hour, but it doesn't change what the board can rely on. Four other things do.
1. Source. Every figure should point back to where it came from: the trial balance, the CIM, the management accounts, the facility agreement. A number without a source is an opinion with decimal places.
2. Separation. Cash is not EBITDA. Adjusted is not reported. A forecast is not an actual. Form is not substance. When these get blended in prose, a reader can't tell what the conclusion actually rests on.
3. Honest gaps. If the quality of earnings report isn't in the file, the record should say so. A tool that quietly fills the missing input with a plausible figure has made the paper less reliable while making it look more complete.
4. A record, not a chat. A board, an auditor, a lender or a buyer will read what you wrote, not the conversation that produced it. The output has to stand on its own.
Where general AI goes wrong in finance
General chat assistants are trained to be helpful, which in practice means producing a complete-looking answer. Give one an LBO question and half a data room, and you'll get a full model. Some of its inputs will be in your file. Others will be typical values for that kind of deal. The answer won't tell you which is which.
That isn't a bug you can prompt away. It comes from asking a general system to behave like a careful finance professional without giving it the professional's rules. CFO Sentinel builds those rules in:
- A number is treated as a fact only when it is in the file, with the source named. Nothing is invented to fill a cell.
- Reasoning from the file is marked as reasoning.
- Missing inputs stay missing. The desk returns a framework and a required-data list, not a fabricated model.
How CFO Sentinel works
You open a matter and attach what you have: the model, the board paper, the trial balance, or the note you have to answer. The desk accepts PDF, Word, Excel, CSV, text and image files. Then you choose one of three intents.
Ask is for a bounded question. Is this a cash problem or an EBITDA problem? How much covenant headroom is there on the facility definitions? What will the buyer contest? In Conservative, Ask stays inside the file. In Research, it may name relevant frameworks, and those are marked as inference.
Review reads the file and its exhibits. Classifications, capital implications, jurisdiction, risk and confidence each sit on their own line. Use it when a pack has to be ranked: which issues matter, in what order, and how sure you can be.
Deep Review is for the paper that leaves the room: the board report, the IC paper, the irreversible call. It runs two independent workings, then settles them into one answer. Where they differ, it decides, and any doubt that remains is stated once. The output runs from executive summary through next steps, as a record that can sit in a pack.
A worked example (fictional file)
Northwind Holdco, PE/LBO, Delaware. Question: can the holdco support a 6.5× debt / LTM EBITDA take-out without a twelve-month liquidity event?
- Evidence. LTM EBITDA of $48m is in the CIM pack. Cash of $11m is on the latest trial balance. Both sit in the file. Neither is a forecast.
- Inference. Covenant headroom looks thin if working capital is seasonal. That is a reading, not a fact in the pack.
- Gap. There is no borrowing-base certificate, no quality of earnings and no revolver draw schedule. The desk will not invent them.
- Record. Risk: High. Confidence: Requires local validation. The next step is the missing data, not a prettier model.
A general assistant would likely have produced a debt schedule. This record tells the deal team what to go and get.
What it covers
FP&A, valuation, M&A, PE and LBO, venture, IPO, accounting, tax and structuring, distress, capital structure and strategy. It is for founders and CFOs, boards and investment committees, PE and VC firms, family offices and corporate development teams.
What it is not
It is not a trading terminal, not retail stock tips and not live brokerage. It is not an audit, tax opinion, legal opinion or regulated investment recommendation. Statute-level conclusions are marked Requires Local Validation. The desk isn't a licensed auditor, tax agent, solicitor or regulated adviser, and it won't borrow a badge it doesn't have.
Getting started
Create an account and confirm your email, and 150 free credits land with no card. Work is paid in credits. Seats are Operator (1,500 credits included a month), Counsel (4,000) and Chambers (12,000). The FIA Plan house seat gives 1,000 credits a period across all seven FIA Labs desks.
Questions
Can AI replace a CFO?
No. CFO Sentinel is decision support. It prepares a record from the file, and a qualified person decides and signs.
Will CFO Sentinel invent numbers to complete a model?
No. A number is treated as a fact only when it is in the file. Missing inputs stay missing, and the desk returns a framework and a required-data list instead.
What files can I attach?
PDF, Word, Excel (.xlsx and .xls), CSV, text and common image formats.
What does it cost to try?
A short Ask usually costs about 1 credit and a short Review about 4. A long, worked answer costs more: around 6 credits for an Ask and 7 for a Review. Before a run, the desk holds enough credits to cover it, and you are charged only for what the run used. A stopped run is not charged. New accounts get 150 free credits with no card.