AI-Assisted Financial Statement Analysis

Analysing financial statements — reading and interpreting the balance sheet, income statement and cash flow statement to understand a business’s financial position and performance — is a core finance skill, and one where AI can genuinely help. A large language model can read a set of accounts, summarise the key figures and movements, calculate ratios and draft commentary, giving the finance professional a fast starting point to develop and verify. What it cannot reliably supply is the understanding and judgement at the heart of the work: interpreting what the statements really show, applying professional scepticism and forming a considered view. AI can also produce analysis that is plausible but wrong, so its help has to be used carefully.

This guide is written for finance professionals who want to use AI to assist with financial statement analysis. It covers where AI helps, why the understanding and judgement must stay with you, how to use AI for the parts it suits, a worked ratio example with figures you can check, how to read the notes and accounting policies, how to work from Companies House filings, how to check AI arithmetic in Excel, the red flags worth looking for and the limitations you need to respect. It sits alongside our wider guides on where AI helps and where it is dangerous in finance and how large language models work.

Where AI Can Help With Analysing Financial Statements

AI can genuinely help with the parts of financial statement analysis that suit its capabilities. It is good at reading and summarising, so it can work through a set of accounts — the primary statements, the strategic report, the notes — and pull out the key figures, the main movements and the points that merit closer attention. For a long annual report, or a batch of accounts to review, that first pass can save real time.

AI can also help compute and present the analysis — calculating ratios, identifying trends, comparing periods or entities — though every figure must be verified. It can draft the narrative around the analysis in clear prose, and it can support the process by suggesting matters to consider and helping structure the work. Reading and summarising, computing and presenting, drafting and supporting: those are the parts where AI accelerates statement analysis, while the interpretation and judgement remain yours. In practice, finance teams use it for four recurring jobs.

Your own company’s accounts

For your own business, AI is most useful for comparing this year’s draft statements with last year’s, listing the movements above a threshold you set, and drafting first-cut explanations for you to correct. It can also review a draft annual report for consistency between the strategic report, the primary statements and the notes. Unpublished numbers are confidential, so use only tools your firm has approved — see our guide to data security when using AI in finance.

Competitors and benchmarking

Published accounts of competitors are public, which makes them a good, low-risk use case. AI can extract margins, revenue growth, headcount and staff costs from several sets of accounts into a consistent table, flag differences in year ends and accounting frameworks, and summarise what each strategic report says about trading. Benchmarking is only as good as its comparability, so check that you are comparing like with like before drawing conclusions.

Customers and suppliers: credit risk

Credit control teams and financial controllers increasingly use AI to triage the accounts of new customers and key suppliers: summarising the balance sheet, net current assets, borrowings and charges, and listing anything that suggests strain — late filing, a going concern note, a qualified audit opinion or shrinking net assets. AI cannot set a credit limit for you, but it can make the review faster and more consistent across a large ledger.

Acquisition targets

In early-stage M&A work, AI can turn several years of a target’s filed accounts into a trend table, highlight changes in accounting policy and related-party transactions, and draft a list of questions for management or for financial due diligence. Information received under an NDA is confidential, and if a listed company is involved it may be inside information, so the data-handling rules matter even more here. For the financial controller’s wider role in deals, see the FC’s role in fundraising and due diligence.

Why the Understanding and Judgement Must Remain With the Finance Professional

Financial statement analysis is not just reading the numbers but interpreting them — understanding what they show about the business, seeing beyond the figures to the underlying reality, applying professional scepticism and forming a considered view. That interpretation rests on accounting knowledge and business insight which AI cannot reliably supply. AI can summarise and compute; deciding what the statements mean is the finance professional’s job.

This matters because AI can produce analysis that sounds insightful but is shallow or wrong: observations that miss what really matters, plausible but incorrect interpretations, or a reading that lacks scepticism. Financial statements can conceal as well as reveal. Aggressive revenue recognition, capitalised costs, a one-off disposal gain or a quiet change in a provision can all flatter a set of numbers, and a model generating plausible commentary may simply describe the flattering version. Relying on AI’s analysis uncritically risks exactly the shallow reading that genuine analysis exists to prevent. The interpretation and judgement are the heart of the work and must remain yours, with AI assisting the reading and computation around them.

How to Use AI Well for the Parts It Suits

Using AI well means letting it handle the reading, computation and first-draft narrative while you supply the interpretation, judgement and verification. Use it to accelerate the initial reading and to help compute ratios and trends, then bring your own understanding to what the statements really show. Treat its summaries, calculations and observations as a starting point to verify and build on, not as the finished analysis. A workflow that works well in practice:

• Extract first, analyse second. Ask the AI to transcribe the primary statements into a table with page references, and check that table against the source before asking for any analysis.
• Define the ratios yourself. State the formula for each ratio (for example, trade receivables at year end divided by revenue, times 365) so the output follows your conventions rather than the model’s.
• Recalculate in Excel. Rebuild the key ratios in a spreadsheet from the verified table — the AI’s numbers are a cross-check, not the record.
• Ask for questions, not conclusions. “What would a sceptical credit analyst ask about these accounts?” tends to be more useful than “Is this company healthy?”
• Write the conclusion yourself. Use AI to tidy the prose, but form and own the view.

Good prompts make a large difference here; our finance prompt library and guide to prompt engineering for finance cover the technique, so this page concentrates on the analysis itself.

A Worked Example: Ratio Analysis With AI

The figures below are illustrative — a fictional UK distribution business — but they add up, so you can recalculate every ratio yourself. Imagine you have asked an AI assistant to extract these items from two years of filed accounts and calculate the standard ratios. Corporation tax is assumed at 25%.

What the ratios say

A typical AI summary would say: “Revenue grew 20% but margins declined and leverage increased.” True, but shallow. The more useful reading is that growth has been bought with working capital and debt. Trade receivables rose 44% against 20% revenue growth; receivable days lengthened by more than 13 days; inventory days rose while payable days fell, so the business is funding customers and stock while paying suppliers faster. Trade working capital (receivables plus inventory less payables) went from £3.7m to £5.9m, an increase of £2.2m — which consumes most of the year’s £3.2m EBITDA before tax, interest or capital expenditure. Cash fell even though borrowings doubled, and interest cover halved.

The questions that follow are the analyst’s, not the model’s: are new customers on longer terms or slower to pay? Is there a concentration of overdue debt? Is the margin squeeze from pricing to win volume? Are the bank facilities committed, and how close are the covenants? AI can help draft these questions once you have seen the pattern, but recognising the pattern is the skill. For more on following working capital through to cash, see building a 13-week cash flow forecast and building a three-statement model.

Where the AI answer could differ — without being wrong

If the AI reports Year 2 receivable days as 66.9 rather than 79.1, it has probably used average receivables ((5,200 + 3,600) / 2 = 4,400) instead of the year-end balance. Neither is wrong, but mixing conventions across companies or periods makes a comparison meaningless. Likewise, “EBITDA” can be stated before or after exceptional items, and “net debt” may or may not include leases. This is why you define each ratio in the prompt and recalculate in a spreadsheet.

Checking AI Arithmetic in Excel

Language models generate text; they do not calculate the way a spreadsheet does. Many tools can now run code or call a calculator, which improves arithmetic considerably, but you should still treat any figure in a chat answer as unverified until a spreadsheet agrees with it. A simple routine:

• Paste the extracted figures into Excel, not the ratios. Build the ratios with formulas, for example =B10/B2*365 for receivable days, so every number traces to a source cell.
• Tie the extract to the accounts. Check that gross profit, operating profit and profit before tax recompute from their components, and that the balance sheet balances, before you trust any line.
• Compare the AI’s ratios with yours. Use a difference column. Any gap is either a convention difference or an error, and you need to know which.
• Watch units and signs. Accounts in £000 or £m, and costs shown in brackets, are frequent sources of slips when AI reads a PDF.
• Check the columns. A model can swap current and prior year, especially where a restatement adds a third column.

If your team uses AI inside Excel itself, our guide to using AI inside Excel for finance covers the practicalities and the controls.

Reading the Notes and Accounting Policies

The primary statements tell you what happened; the notes and accounting policies tell you how the numbers were made. This is where AI’s ability to read long documents quickly is genuinely valuable — and where its tendency to summarise rather than question is most dangerous. Useful things to ask an AI to find, with page references, include:

• Revenue recognition. When revenue is recognised, how multi-element contracts are handled, and any judgement about principal versus agent.
• Changes in policy or estimates. Any change in depreciation lives, provisioning, capitalisation of development costs or stock valuation — and the effect on profit.
• Prior-year restatements. Restated comparatives can signal errors or policy changes that alter the trend.
• Critical judgements and estimation uncertainty. Where management says the numbers depend most on judgement.
• Going concern. The directors’ assessment, and any material uncertainty disclosed by the directors or highlighted by the auditor.
• Borrowings, security and covenants, commitments, contingent liabilities, related-party transactions, directors’ loans and post-balance-sheet events.

Comparability needs particular care at the moment. The FRC’s periodic review of FRS 102 introduced a five-step revenue model and brought most leases onto the balance sheet for periods beginning on or after 1 January 2026. For a lessee, rent that used to sit in operating costs is replaced by depreciation and interest, so EBITDA and reported net debt can both rise without any change in the underlying business. An AI comparing a pre-2026 year with a post-2026 year will not necessarily notice, so check the transition note. The Financial Reporting Council publishes the standards themselves.

Working From Companies House Filings

For competitors, customers, suppliers and private acquisition targets, the starting point is usually the free Companies House register. As well as filed accounts, the filing history shows confirmation statements, changes of directors, registered charges (security granted to lenders), the register of people with significant control and any overdue filings — all useful context that an AI can summarise if you give it the documents. Private companies normally have nine months from their year end to file accounts, according to GOV.UK guidance, so filed figures can be well over a year old by the time you read them.

Know the limits of what has been filed. Many small companies currently file abridged or “filleted” accounts without a profit and loss account, and micro-entities file very little. From 6 April 2025 a company is small if it meets two of: turnover up to £15m, balance sheet total up to £7.5m and up to 50 employees (medium: £54m, £27m and 250). Reforms under the Economic Crime and Corporate Transparency Act will require small and micro companies to file a profit and loss account; that change was paused from its original April 2027 date and is now expected from April 2028, with an option to keep it off the public record still to be confirmed — ICAEW tracks the latest position. Until then, an AI asked to analyse a small company’s margins may have no income statement to work from, and you should be wary of any answer that supplies one.

Most accounts are filed in iXBRL, a tagged format that makes figures machine-readable, and Companies House offers bulk data and an API. Feeding tagged figures into a spreadsheet is generally more reliable than asking an AI to read a scanned PDF, which may be a poor-quality image of a printed document.

The Safeguards Required

The analysis informs understanding and decisions — credit limits, supplier choices, bids, board papers — and AI can produce plausible but wrong output, so safeguards are essential. The foundational one is that the finance professional supplies the interpretation and judgement, using AI to assist the reading and computation but not to determine the analysis. Ownership of the interpretation is the essential safeguard.

The related safeguards are: verify AI’s contributions — figures, summaries and observations — against the source; apply professional scepticism to both the statements and the AI’s reading of them; do not rely on AI for the judgement-intensive interpretation it cannot reliably do; protect the data, using only approved tools for confidential statements and following your firm’s AI usage policy; and keep a record of what was AI-assisted and what was checked, which is the kind of evidence covered in what auditors ask about AI. A finance professional who applies these safeguards uses AI for statement analysis safely; one who does not risks a shallow or wrong analysis.

Red Flags Worth Asking AI to Look For

A checklist prompt is one of the most effective uses of AI in this work: it is quick, consistent and makes the model look for specific evidence rather than offer a general opinion. Ask it to report each item as found, not found or not determinable, with a page reference, then check the “found” items yourself.

• Receivables or inventory growing materially faster than revenue.
• Profit rising while operating cash flow falls, or a persistent gap between the two.
• Recurring “exceptional” or “non-recurring” costs, or one-off gains propping up profit.
• Changes in accounting policies, estimates or year end, and restated comparatives.
• A modified audit opinion, a material uncertainty over going concern, or a change of auditor.
• Late filing, extended accounting periods or frequent changes of directors.
• New charges registered, rising borrowings, falling interest cover or covenant waivers.
• Significant related-party transactions or directors’ loan balances.
• Growing capitalised costs, such as development spend, relative to revenue.
• Net current liabilities or net liabilities, or a sharp fall in net assets.

None of these proves a problem; each is a reason to look harder. For context on how often businesses fail, the Insolvency Service publishes company insolvency statistics, which is useful background when setting credit policy.

How to Approach AI-Assisted Statement Analysis Sensibly

Approaching AI-assisted analysis sensibly means using AI to accelerate the reading, assist the computation and draft the narrative, while you supply the interpretation, apply scepticism and verify. It means applying the safeguards so that AI assists the analysis rather than determining an unsound reading, and treating AI as a tool that supports the work, not a replacement for the understanding and judgement it requires.

It also means using AI’s output critically — checking its figures, questioning its observations and bringing the same scepticism to its analysis as to the statements themselves. The value of financial statement analysis lies in the understanding it produces, and that requires your interpretation, not just a model’s summary. A finance professional who works this way captures the time saving while keeping the analysis sound; one who relies on AI uncritically risks a shallow or wrong reading. This connects to the guidance on human-in-the-loop AI controls and to our grounded view of what is real and what is hype in AI.

Where AI Analysis Is Most and Least Reliable

It helps to distinguish where AI’s assistance is more reliable and where it is less, so you can judge how much to lean on it. AI tends to be more reliable on the mechanical and factual aspects — reading and summarising what the statements say, computing ratios from given figures, identifying movements between periods — because its output can be checked against the source. Verification is still needed, but the output is checkable and the task suits the tool.

AI tends to be less reliable on the interpretive aspects — what the numbers really mean, what might be concealed, whether management’s narrative fits the figures, what the business’s position really is — because these need understanding and judgement it cannot supply, and here its plausible output is most likely to be shallow or wrong. Lean on AI for the mechanical and factual work, rely on your own understanding for the interpretation, and treat its interpretive comments with particular caution.

Hallucination and Other Limitations

“Hallucination” is the term for a model producing confident content that is not supported by its source or by fact. In statement analysis it shows up in specific, recognisable ways:

• Invented figures — a profit and loss line for a company that filed only a balance sheet, or a prior-year number that is not in the document.
• Invented references — citing “note 14” when the relevant note is note 17, or quoting wording that does not appear.
• Misread tables — swapped columns, lost brackets, or £000 read as pounds.
• Background leakage — mixing what it “knows” about a well-known company from training data with what is actually in the accounts in front of it.
• Out-of-date knowledge — stating thresholds, tax rates or accounting rules as they stood when the model was trained.

Practical defences are to give the model the document rather than asking from memory, instruct it to answer only from the document and to say when something is not there, require page references, and verify anything that matters. Our guide to large language models for finance professionals explains why these errors happen, and the ICAEW’s artificial intelligence resources cover the professional and ethical considerations for members.

Using AI to Deepen Rather Than Replace Analysis

One of the most valuable ways to use AI in statement analysis is to deepen the analysis rather than replace it. When AI handles the mechanical work — reading, extracting, summarising, first-pass computation — your time is freed for interpretation, scepticism and judgement. Used that way, AI does more than speed the work up: it lets you review five years instead of two, benchmark against eight competitors instead of three, or read every note rather than skim.

The contrast is using AI to replace the analysis — accepting its output as the finished product — which risks the shallow or wrong reading its plausible generation may produce. A finance professional who uses AI to free attention for interpretation gets more from both the AI and their own judgement. That is the most valuable way to use AI in statement analysis, and it enhances rather than diminishes the understanding the analysis should produce.

What This Means When You Are Hiring

For employers, the skill worth paying for is not the ability to prompt a chatbot but the ability to read accounts critically and use AI to cover more ground. In interviews for FP&A roles, management accountants or financial controllers, a useful exercise is to hand candidates a set of filed accounts and an AI-generated summary containing a deliberate error, and see whether they find it. Our article on how AI is changing what employers look for in an FC covers the wider shift. Where the analysis feeds board-level decisions or a CFO appointment, our sister firm FD Capital handles finance director and CFO searches.

Frequently Asked Questions

Can AI analyse a set of financial statements on its own?

It can produce a fluent summary and calculate ratios, but it cannot be relied on to interpret them soundly or to spot what the accounts may be concealing. Use it for extraction, calculation and first drafts; the conclusion should be yours, verified against the source.

Is it safe to upload accounts to an AI tool?

Accounts filed at Companies House are public, so the confidentiality risk is low, although your firm’s policy still applies. Your own unpublished figures, management accounts and anything received under an NDA are confidential and should only go into tools your organisation has approved for that data.

Why do AI ratios sometimes differ from mine?

Usually because of definitions — year-end versus average balances, EBITDA before or after exceptional items, net debt with or without leases — and sometimes because of a misread figure. Define the formula in your prompt and recalculate in Excel so you know which it is.

Can AI assess the credit risk of a customer?

It can speed up the review by summarising filed accounts and flagging warning signs, but small companies often file limited information, the figures may be well over a year old, and AI cannot weigh payment history or trading relationships. Use it as one input alongside credit reference data and your own ledger experience.

Will the FRS 102 changes affect year-on-year analysis?

Yes, potentially. For periods beginning on or after 1 January 2026, most leases come onto the balance sheet and revenue follows a five-step model, so EBITDA, debt and the timing of revenue can move without any change in trading. Read the transition disclosures before comparing periods.

Which AI tool is best for financial statement analysis?

It depends on what your organisation has approved and how the tool handles documents and calculations. Our comparison of Claude, Copilot and ChatGPT for finance sets out the differences; the verification steps on this page apply whichever you use.