How to Use AI to Run Finance Like a Top CFO

Finance teams often ask AI to summarize last month's results. A faster summary is useful, but it does not recover the decision window that has already closed. The more valuable question is: What is changing now, what might explain it, and which choice can we still make?

That is the discipline behind my book, The Impact of AI in Financial Decisions. AI can organize evidence, calculate scenarios, and challenge a recommendation. It cannot be allowed to turn a plausible explanation into an approved financial action on its own. A strong CFO uses the tool to shorten the distance between a signal and a well-supported human decision.

The margin beat that hid a gross-profit miss

Chapter 5 opens with a finance team delivering a prior-month variance analysis on day 12 of the following month. By then, the business is already halfway through another trading period. The book later gives an illustrative retail example: budgeted revenue of $50 million at a 40% gross margin, versus actual revenue of $46 million at a 42% gross margin. The margin rate beat plan by two percentage points, but gross profit was $19.32 million rather than the budgeted $20 million—a $680,000 shortfall. Those are illustrative book figures, not a result promised to any business.

The percentage alone could invite congratulations; the revenue miss alone could invite a blanket discount. Neither reaction is a decision. The book's scenario also mentions weaker store traffic, canceled promotions, and growing online sales. A finance leader needs to verify how much each driver contributed, whether it persists, and what the proposed response would do to contribution margin and cash. AI can prepare that investigation while the next promotion or inventory allocation is still open to change.

1. Start with the decision, not the dashboard

Write the decision in one sentence: “Do we change the next two weeks' promotional plan, reallocate stock, or hold course while we test demand?” Add a deadline, an owner, and the constraints. A dashboard tells you what happened; this question tells the finance team what evidence it needs before a choice is made.

Give an approved AI tool a small, controlled pack: comparable actuals and budget, units and average selling price, product and channel mix, discounts, inventory, and the assumptions behind the current forecast. Label the source and period for every file. Keep customer-sensitive data out of unapproved tools. Ask the model to identify missing information before it offers an answer.

2. Separate the movement from the story

The book's variance exercise classifies movements as timing, volume, price or rate, and mix. Apply that discipline to the revenue gap. Did fewer customers buy, did they buy fewer units, did realized prices fall, or did a shift toward different products change the total? A delayed order is not the same as a lost customer, even if both depress this month's revenue.

Have AI produce two columns: “verified from the supplied records” and “possible explanation requiring confirmation.” Require it to show the arithmetic and the source rows. Finance should recalculate the totals independently and ask sales, operations, and customer-facing colleagues what the records cannot show. If periods, definitions, or source extracts are not comparable, stop the analysis and fix that first.

3. Make the preferred response defend itself

Suppose the instinctive answer is a broad discount. Ask AI to build the strongest case against it. How many extra units would be needed to recover the gross profit sacrificed per sale? Would it pull forward orders that would have arrived anyway? Could a targeted offer, a stock transfer, a sales follow-up, or no intervention produce a better outcome? Which assumption would reverse the recommendation?

This is where the top-CFO habit matters: use AI while the spending, pricing, or allocation choice is still reversible. Do not ask the model to “pick the winner” from invented inputs. Give it approved numbers, request a base case and downside case, and make it show the break-even condition. Finance and the business owner then decide whether a small, measurable test is worth running.

4. Bring cash into the same conversation

An attractive profit story can still hide a cash problem. The book's treasury discussion points to receivables aging, payables schedules, payroll cycles, and seasonal patterns as inputs to cash forecasting. The rolling-forecast exercise asks the analyst to update the year-end view as new actuals and assumptions arrive. Use those ideas together: if sales move, what happens to collections, inventory cash, supplier payments, and the next forecast?

Ask AI for a short exception list, not a wall of charts: which change is material, when cash could feel it, what evidence supports the estimate, and what must be checked with the people who own the process. The model can draft a scenario; treasury and FP&A must validate the timing and the numbers. A revenue decline that is merely delayed calls for a different response from a persistent loss of demand.

5. Turn analysis into a decision with a checkpoint

End with a one-page decision note: the choice, the alternatives considered, the evidence, the assumptions that matter most, who has approval authority, and the date on which the team will review the result. Tell the team what it may do now and what still requires a CFO or business-owner sign-off. For the retail example, the checkpoint might be the next comparable week of orders, unit margin, promotion uptake, and collections—not a vague promise to “monitor performance.”

AI can draft the note and flag contradictions. A person checks the evidence, speaks with the relevant teams, and owns the final call. That division of work is central to the book: better analysis should strengthen financial judgment, not obscure who is accountable.

A prompt to use with your own numbers

Put one real, current decision into an approved tool. Redact or minimize sensitive information, specify the source of every figure, and ask for a reviewable output. The prompt below is a starting point, not a substitute for reconciled data or a finance professional's approval.

Act as a skeptical FP&A partner. Our decision is [choice] by [date].
Use only the attached approved actuals, budget, forecast, and assumptions.

1. Recalculate the key movements and show the formula and source for each.
2. Classify each variance as timing, volume, price/rate, or mix where supported.
3. Separate verified facts from hypotheses and list the evidence needed to test them.
4. Compare at least three responses, including a small reversible test and no change.
5. Show base and downside effects on gross profit, cash timing, and customer trust.
6. Challenge our preferred option: what assumption could make it the wrong choice?
7. Draft a decision note with an owner, approval boundary, and review date.
Do not invent missing inputs, change records, or approve an action.

Start with one decision this week. If the exercise only gives you a polished recap, narrow the question until it reveals a choice, an uncertainty, and a test. The point is to give your people time to act before the next close turns today's opportunity into last month's explanation.

P.S. Finance teams get more from AI when they use it on decisions that can still change. Book a discovery call with us, and we'll shape the workflow around your real numbers, your approval process, and the outcomes that matter to your business.