How to Use Claude to Do Financial Analysis

Define the decision before opening a statement

Start with one sentence: what decision will this analysis support, who will read it, what period matters, and what action is still reversible? 'Analyse these accounts' is too broad. 'Explain the change in gross margin between the two audited periods and list questions for management' gives Claude a boundary. Add currency, accounting basis, consolidation scope and the date the analysis must be current to.

Tell Claude what it must not do. It should not recommend a security, estimate tax liability, infer fraud, fill a missing number or present a scenario as a forecast. Ask it to label each output OBSERVED, CALCULATED, ASSUMED, INTERPRETED or UNKNOWN. Those labels make the later review much easier.

Prepare a redacted source pack

Use the smallest source set that can answer the question: annual or quarterly reports, notes, management commentary, a budget extract and a definitions sheet. Remove account numbers, personal data, credentials, customer identifiers and unrelated contracts. Keep a manifest with filename, period, version, source owner and whether the file is audited, unaudited, internal or illustrative.

Claude Projects can hold documents and project instructions, but product availability and account controls change. Check the current Anthropic documentation and your organisation's policy before uploading financial material. A project workspace is not an approval system. Preserve the original files separately and give the model a copy or a redacted export.

Ask for an inventory and accounting map

The first response should be a source inventory, not an opinion. Ask Claude to list each file and period it can read, statement type, currency, units, reporting basis, comparative period, page or table location, and any unreadable or contradictory section. Then ask for a mapping of revenue, cost, asset, liability, cash-flow and non-GAAP terms to the exact source wording.

If the same label changes definition across periods, stop and resolve it. Revenue can be gross or net; cash flow can be operating or free cash flow; a margin can use different denominators. Ask Claude to quote a short source fragment or give a page reference for each definition, then open the original yourself. No ratio is meaningful until its numerator, denominator and period are explicit.

Build a reproducible variance analysis

For each material change, require a table with prior value, current value, absolute change, percentage change, unit, source location and a plain-language explanation. If the explanation is not directly supported, label it as a hypothesis and list the evidence needed. Ask for a bridge or waterfall structure only after the arithmetic ties to the reported totals.

Recalculate a sample in Excel or another controlled tool. Check signs, parentheses, currency conversion, discontinued operations, rounding and whether the statements use thousands or millions. Ask Claude to identify values that are extracted from a table versus values it calculated. A 10% change is not an explanation; it is a calculation that still needs an evidence-backed driver.

Use ratios as questions, not verdicts

Claude can calculate and explain ratios such as gross margin, operating margin, current ratio, debt-to-equity or working-capital days when you provide the definitions and inputs. Require the formula, inputs, periods, units and interpretation limits beside every ratio. Avoid comparing ratios across companies or industries without checking accounting policies, capital structure, seasonality and source dates.

Ask for a diagnostic list rather than a verdict: what changed, what could explain it, what evidence would distinguish the explanations, and which question belongs with management or a professional adviser. A ratio can identify where to look; it rarely proves why something happened. Keep a calculation ledger that another analyst can reproduce without access to the chat.

Separate historical analysis from scenarios

Historical statements describe supplied results. A scenario changes assumptions. If you ask Claude to model a downside or upside case, state the starting period, variables, ranges, dependencies, timing and what is held constant. Ask it to return the assumptions table first, then the formula for each output and a sensitivity table showing which assumption matters most.

Do not let words such as likely, expected or conservative appear without an owner and evidence. Call it a scenario, not a forecast, unless a qualified process supports that label. Never treat generated price targets, trading signals or return projections as reliable investment research. The model can structure questions and arithmetic; it cannot know future prices or your risk tolerance.

Ground current facts in primary sources

Financial analysis often mixes internal figures with current prices, interest rates, regulations, tax rules or company announcements. Tell Claude which primary sources are authoritative for each fact: an issuer filing, regulator, central bank, official tax authority, audited report or approved internal system. Require the source date and a link or document location. Open the source before using a number in a decision.

For public-company work, distinguish reported results from management claims and analyst interpretation. For private-company work, distinguish the supplied management number from independently verified evidence. If a source is unavailable, write UNKNOWN rather than asking the model to complete the gap from memory.

Protect sensitive financial data

Financial files can contain personally identifiable information, payroll, banking details, customer balances, commercially sensitive prices and privileged advice. Redact what the question does not need. Use approved accounts, least-privilege sharing and the organisation's retention rules. Do not paste credentials, account numbers or full transaction exports into an unapproved chat.

Ask Claude to list the sources it used and the fields it could not inspect. Check links, file permissions and generated downloads before sharing them. If the analysis affects a customer, employee, borrower, investor or regulated report, involve the responsible owner and record the review. Privacy controls reduce exposure; they do not turn AI output into professional advice.

Run an adversarial review and handoff

After the first analysis, ask for the strongest case that the conclusion is wrong, the three most sensitive assumptions, missing data that could reverse the interpretation, and every number that needs recalculation. Then perform a separate human review of extraction, arithmetic, definitions, source dates, units, rounding and wording. Do not let the same generated answer serve as both analyst and approver.

The handoff should contain the source manifest, calculation ledger, assumptions, unresolved questions, scenario boundaries, reviewer name, approval date and the final version. If the result is for an investment, lending, tax, legal, audit or compliance decision, route it to the qualified professional or established control process. The safest output may be a better question list, not a recommendation.

A worked review loop for a margin question

Suppose the question is why gross margin changed between two quarters. Do not upload the files and ask for an explanation in one step. First provide the statement pages, accounting definitions and whether revenue is gross or net. Ask Claude to return the exact revenue and cost figures, units, periods and page locations. If a value is absent or a note is ambiguous, stop there.

Next ask for the arithmetic only: revenue minus cost, divided by revenue, with the prior-period comparison and a control total. Recalculate it in a spreadsheet. Then ask for a driver table covering price, volume, product mix, input cost, foreign exchange, one-off items and unknowns. Each driver needs a source or must be labelled as a hypothesis. Do not accept “higher costs” without a line item, management note or other evidence. Finally ask for a briefing with verified change, supported drivers, open questions and proposed next action in separate sections.

Use the same loop for liquidity, debt or working-capital questions. Start with definitions, trace inputs, test arithmetic, then ask what evidence would change the interpretation. If the result affects a covenant, tax filing, lending decision, investment, client recommendation or regulated report, treat the output as working-paper support and route the conclusion through the required professional review.

Handle conflicting periods, currencies and definitions

A common failure is a comparison that looks precise but combines incompatible inputs. Ask Claude to make a period table showing fiscal year, quarter, calendar dates, currency, unit scale, consolidation scope, restatements and whether each figure is reported, adjusted or estimated. If one document uses dollars and another uses euros, do not let the model choose an exchange rate from memory. Supply the approved rate, date and source, or leave the comparison unresolved.

Likewise, do not merge EBITDA, operating profit and net income because they are all called earnings in a conversation. Ask for a glossary that preserves the source term and gives a plain-language explanation. When management uses an adjusted measure, show the reconciliation or label the adjustment as unavailable. The analysis becomes more useful when it exposes an apples-to-oranges comparison early. A qualified reviewer can then decide whether a normalisation is appropriate and document the policy behind it.

Write the limits where the reader will see them

Put the important limitation beside the result, not buried in a footnote. Say whether the figures are supplied, unaudited, rounded, restated, incomplete or scenario-based. Name the source date and the next review date. If a ratio depends on a denominator that may change, say so. Clear limits protect the reader from treating a clean table as stronger evidence than it is and help the next analyst know what to verify first. If the evidence is weak, make that the headline rather than hiding it behind a confident recommendation. A visible caveat is part of a useful answer.