AI & financial models

How to check an AI-built financial model before you rely on it

Check the opening balances, cash logic, assumptions and scenarios in an AI-built model before using it for hiring, spending or funding decisions.

Before relying on an AI-built financial model, match its starting figures to your records, trace transactions to cash and test changes in assumptions. Review the calculations and business logic: a workbook can balance while omitting a payment or assuming uncertain funding arrives.

1. Establish what the model is meant to answer

Write down the decision: hiring, spending, funding or a board plan. Confirm the period, currency, units and last actual month. Identify the worksheets and data sources behind the output.

Save a review copy and record its version and assumptions. Check that linked workbooks are the intended files and contain current values.

2. Check the starting point against real records

Match opening cash to reconciled bank balances. Check unpaid customer invoices, supplier bills, payroll liabilities, tax balances and borrowing against the accounting records. Identify transactions after the reporting date that already affect the next period.

A correct bank balance is insufficient if the model forgets bills owed. Collecting an old invoice should not create new revenue. Check how opening balances clear or carry forward.

3. Follow a few transactions all the way through

Choose examples important to your business: a customer paying an annual subscription, a new employee starting partway through a month, an equipment purchase and a VAT settlement. Write down the expected accounting and payment sequence before inspecting the formulas.

In a model with linked financial statements, check the effect on profit, amounts owed or held, and cash. Payroll should reflect employer costs and payment dates without duplicating deductions. Tax and pension assumptions should come from the company’s circumstances and current records.

For a cash-only forecast, check receipts and payments directly and label that narrower scope. Do not infer that profit or the balance sheet has been tested when the workbook does not model them.

4. Challenge assumptions and unexplained balancing figures

Find the main assumptions: conversion rates, prices, customer losses, hiring dates, margins and payment delays. Ask who supplied them and what supports them. Distinguish agreed contracts, management estimates and unresolved gaps.

Look closely at rows called ‘other cash’, ‘adjustment’, ‘funding required’ or similar. The name alone does not establish an error. The issue is whether a value has been inserted simply to make a balance sheet balance or keep cash positive, without an explained transaction behind it.

Planned equity funding can be a legitimate scenario assumption, but it should have an explicit amount and date and remain distinguishable from committed cash. A loan facility should reflect its actual availability, limits and terms. A calculated funding gap is an output requiring action; it does not mean funding has been secured.

5. Test whether the model behaves as expected

Predict the direction of the result before changing an input. Delay a customer’s payment: cash should move later and, where modelled, the amount owed should remain outstanding longer. Push a hire back: the related employment costs and payments should move too.

Try a slower-sales case, a delayed raise and one assumption at zero where that is meaningful. Check the whole model, including the final forecast period. Do not fix an unexpected result by overwriting the output.

The ICAEW Financial Modelling Code recommends comparing model behaviour with expectations, building visible checks and using peer review when the model’s impact is significant. A passing balance check confirms one relationship; it does not confirm that every assumption is realistic.

6. Make the workings reviewable

Check that assumptions are separate from formulas, repeated calculations are consistent and important workings are accessible. Unexplained hard-coded numbers inside formulas make changes harder to assess. Clear ownership, documentation and version control are also part of ICAEW’s good spreadsheet practice.

Recalculate in your spreadsheet application. Investigate broken links, formula errors, unexpected pasted values and circular references. Deliberate circular calculations need a documented purpose and controls.

Ask AI to flag unusual formulas, then verify its findings. Microsoft’s Copilot guidance notes that results can be inaccurate. An automated ‘all checks passed’ message does not establish what was tested.

A review checklist before making a commitment

  • The question, forecast period and latest actual figures are clear.

  • Opening balances agree to the records and unpaid items are included.

  • Important transactions reach the right periods and cash movements.

  • Assumptions have an owner and uncertain funding is visible.

  • No unexplained figure conceals a cash shortfall or balance error.

  • Scenario tests behave as expected and the results have been reviewed.

  • Known limitations and the next update date are recorded.

Frequently asked questions

Can another AI agent perform the review?

It can provide another set of checks, but may share the same misunderstandings or miss context. For a material decision, include a person who understands the business and can challenge the financial logic.

Should we repair the workbook or rebuild it?

Repair may suit a clear model with limited defects. A rebuild may be more useful when sources, assumptions and formulas cannot be traced. Agree the review scope first, then decide based on what is found.

Choose help around the job you need done

To identify broader priorities, get a free Founder finance review: a scored checklist of your finance setup and clear next steps. Detailed model review, repair or rebuild work is paid separately.

Reporting & planning provides monthly or quarterly reports, forecasts and review. Finance leadership (fractional CFO support) adds ongoing help with decisions through a flexible monthly retainer. Both require our Accounting package. For a new draft, see our AI budget and cash forecast guide.

ABOUT STARTUP CFO

Led by Ryan Thomson CA

Chartered Accountant with more than 10 years of experience

Ryan is a Chartered Accountant with 10+ years of experience supporting early-stage startups, with particular expertise in deep tech. He also leads finance at Post Urban Ventures, a London deep-tech venture studio.

ABOUT STARTUP CFO

Led by Ryan Thomson CA

Chartered Accountant with more than 10 years of experience

Ryan is a Chartered Accountant with 10+ years of experience supporting early-stage startups, with particular expertise in deep tech. He also leads finance at Post Urban Ventures, a London deep-tech venture studio.