This website uses cookies

Read our Privacy policy and Terms of use for more information.

Budgeting has been part of my fractional CFO work for years. One client has hired me only for their budget since 2023, which makes it a good place to see how much the process has changed.

In 2023 the budget took me up to 60 hours. Most of it went into collecting data, building the file with formulas and scenarios, and validating the calculations.

In 2024 I used AI to analyze the prior year and organize the data; the work itself was still mostly manual, in spreadsheets.

In 2025 Claude built the file. The result was cleaner, faster and deeper: detailed inputs for every department, built-in dashboards and a variance tracker.

In 2026 the CEO built the budget with Claude and sent it to me with a request to review it.

I knew this day would come. A budget, a variance analysis or any other complex file is now within reach of most people, and it takes no advanced prompting.

So the question changes: if the file takes a few hours, what does the CFO do with the rest of the time? Check it, challenge it, stress test it, and look for the growth opportunities it leaves out. In my client work, that last question almost never gets asked.

This works on a budget you built yourself, or received from your finance team. Claude saved you the build hours; put them to work.

Today I'll show you how to review an AI-built budget in five steps, and set it up with three Claude skills you can download and a short prompt for each step.

How to review an AI-built budget in Claude

TIME: 2 to 5 hours

YOU NEED: budget, actuals, assumptions list

YOU GET: findings, stress test, growth questions

The skills are built for Claude. Download each one from the link in its step, add it to Claude, then run the short prompt under it.

Five steps to review an AI-built budget: collect the budget and its assumptions; check the math; break the model into drivers and challenge each assumption; stress test the drivers; find growth the budget leaves out. Steps 3 to 5 are marked as where the CFO's hours go now.
STEP 1

Collect the budget and the assumptions behind it

Put the budget file, last year's actuals and the prior budget in one Claude Project. Then ask whoever built the budget, you included, to list every assumption they made along the way. The file shows the arithmetic; the list shows the thinking. It also gives Claude something to compare against when it looks for assumptions hidden in the formulas.

STEP 2

Check the math before anything else

Totals tie, rollovers between tabs match, twelve months add up to the year, and formulas are still formulas. A budget built with AI can be very good; after many rounds of iteration it can also carry broken formulas, pasted values and leftover rows. This work is deterministic, so Claude runs it as code: turn on code execution and file creation in Claude's settings before you start.

The budget-integrity-check skill runs five families of checks: math, formulas, logic, missing data and leftovers. It reports only and leaves the file as it was, and it lists the checks that passed so you can see what was tested.

When I ran it on a real client budget, it returned more than 20 findings. Some were serious; others were housekeeping. Without this check, the budget would have gone out inaccurate. Reviewing every finding took me between 30 minutes and an hour.

Download: budget-integrity-check.skill runs the five check families and reports without editing the file.

Run it: a new chat inside the Claude Project that holds the budget.

/budget-integrity-check on [BUDGET FILE].
Compare against [PRIOR YEAR ACTUALS].
Report only; do not change the file.
STEP 3

Break the model into drivers and challenge every assumption

Drivers are the few inputs that move revenue, cost and cash the most: volume, price, headcount, timing. Claude lists them, finds every assumption behind each one, including the ones that sit in formulas and never made the builder's list, and rates each against last year's actuals as supported, weak or no support. For every weak one you get a specific question to take to the function leader who owns the number.

A weak rating means the assumption breaks with history. The business may really be changing, which is why the output ends in a question and you make the call.

Download: budget-driver-review.skill builds the drivers table, rates every assumption, then runs the stress test in Step 4.

Run it: the same Project, a new chat.

/budget-driver-review Part A on [BUDGET FILE].
Compare against [PRIOR YEAR ACTUALS] and [ASSUMPTIONS LIST].
Save the drivers table as a file when you finish.
STEP 4

Stress test the drivers that matter

Claude ranks the drivers by how much a move changes profit and cash, then runs a downside and an upside case for the top three, sized from the range in last year's actuals. Each case comes with a trigger, the signal you can watch that tells you the case is happening, and the decision you would take. The budget becomes a plan the CEO can steer by during the year.

Run it: the same chat, once you have reviewed the drivers table.

/budget-driver-review Part B
Minimum cash balance: [CASH FLOOR].
Stress test the top [NUMBER] drivers.
STEP 5

Look for the growth the budget leaves out

First, use Research mode on public information about the company to build a sourced SWOT, or start from the one the company already has. Keep the budget file out of that chat. Then, back in the Project, Claude combines the research with the drivers table and looks for opportunities the budget does not include: pricing and mix, customers, cost structure, timing, and what the research surfaced.

Every opportunity comes with the cheapest test to run first and what would prove it wrong. These are hypotheses for a conversation with the CEO, and you check every research finding against its source before it gets there. This is the step that turns a review into strategic work.

On my budget, the skill came back with five to nine opportunities, in pricing, product mix and upsells to existing customers. As the CFO, you may not have the latest picture of every customer or product line, so treat the list as an agenda for a conversation with the leadership team.

Download: budget-growth-finder.skill builds a sourced SWOT in Research mode, then lists growth the budget leaves out, each with a test.

Run it: first, a new chat outside the Project with Research mode turned on.

/budget-growth-finder Part A
Company: [PUBLIC DESCRIPTION], [INDUSTRY], [REGION], revenue about [RANGE].
Horizon: next [NUMBER] months. Budget drivers: [DRIVERS LIST].
SWOT: [ATTACHED OR NONE].

Run it: then, back in the Project, with the research report added.

/budget-growth-finder Part B with [DRIVERS TABLE] and [RESEARCH REPORT].

Where the hours go now

Running all five steps took me 2 to 5 hours, depending on how deep I went. For a fractional CFO, those are the hours that used to go into building the file; for a full-time CFO, they come out of the days the build used to take. Either way, the time moves from building formulas to asking the questions the business needs answered.

CHECK BEFORE YOU TRUST IT

  • Totals tie. Pick two totals from the integrity report and recalculate them yourself in the workbook.

  • Drivers point to real cells. Open three cells from the drivers table and confirm the number matches the budget.

  • Research is sourced. Open the source link and date for each of the five findings marked most likely to change the budget.

WHERE IT BREAKS

  • The integrity check returns a long list. More than 20 items came back on my budget, serious ones mixed with housekeeping. Sort by severity and fix the HIGH items before you move to Step 3.

  • Growth ideas only know what is in the files. You may not have the latest view of pricing or customers either; take each idea to the leadership team as a question, with its test attached.

  • Stress test cases look precise and rest on one year of history. Give your own range for any driver where last year was unusual.

Want this built for your team?

I train finance teams on Claude using their own close, reporting and forecasting work.

Stuck on an AI task in your finance work, or want to reach me? Just reply to this email; the most common questions shape the next editions.

And if this was useful, forward it to a colleague.

Anna Tiomina
CFO & AI Trainer