Prompts for financial analysis that tie out to the source

A model reads "Revenue 4,512" in a 10-K table. The header says "in millions"; the line above it is the prior year. Getting the right figure means getting the column, the unit and the period right at once, and that's where AI finance answers fail.

The best-known test of this is FinanceBench, built from public-company filings. With a standard retrieval set-up, GPT-4-Turbo answered only 19% of a 150-question sample correctly; 13% were wrong and 68% were refusals. Handed the exact page that held the answer, it still got 15% wrong (Islam et al., FinanceBench, 2023). That's an older model, and newer ones do better. But the failure pattern hasn't changed: the hard part is reading the right number, not doing the sum.

Finance teams feel it. In a Datarails survey of 270 US finance leaders, 65% said their most common frustration with AI was "confident answers built on the wrong data", and only 5% trusted it to produce board-ready reports without human review (Datarails, October 2026; a vendor survey, so weigh it accordingly).

The prompts below are built around four rules. Pin down source, period, units and sign before calculating. Show every calculation. Keep reported, derived and assumed figures visibly apart. And make "not in the documents" an acceptable answer.

Put this in your Claude or ChatGPT project instructions so it applies to every chat, or paste it above any one-off finance question.

Rules for all financial analysis in this project:

1. Figures: take every number from the attached documents. Quote the line item, the period and the page or tab it came from. If a figure isn't there, write NOT IN SOURCE and do not estimate it.
2. Units: state the unit and currency for every figure (e.g. USD millions). Convert only when asked, and show the conversion.
3. Periods: label fiscal periods as the company does (e.g. FY2026 Q2) and say when a company's fiscal year differs from the calendar year.
4. Signs: show costs as positive numbers unless told otherwise; call a variance favourable or adverse, not just up or down.
5. Calculations: do all arithmetic with the code tool. Show the formula and the inputs for every derived figure.
6. Labels: mark each figure [REPORTED], [DERIVED] or [ASSUMPTION].
7. Definitions: for any ratio or metric with more than one common definition, state the one you used.

Each line closes a specific gap. Rule 1 is Anthropic's and OpenAI's grounding advice applied to numbers, with the out spelled out; OpenAI's GPT-5.5 guide asks models not to "invent specific names, first-party data claims, metrics" and to use "clearly labeled assumptions" instead. Rule 2 catches the thousands-vs-millions slip. Rule 3 catches the company whose FY2026 ended in June. Rule 4 stops "costs went up 8%" being read as good news. Rule 5 follows Google's advice to use code execution "whenever the model needs to perform any kind of arithmetic", and rule 6 is what makes the answer auditable at a glance.

The preamble asks the model to label its assumptions; Count shows whether it did. Connected to Claude, it tags each claim as a citation, an assumption, model knowledge or a hallucination.

From the attached [10-K / annual report / board pack], extract the following for [FY2025] and [FY2026]: revenue, gross profit, operating income, net income, cash from operations, capital expenditure.

First, in <quotes>, copy the exact table row or sentence for each figure, with its page number and the table's unit heading.
Then give a table: line item | FY2025 | FY2026 | unit | page.
If the document reports a figure under a different name (e.g. "income from operations"), say which name it used. If it reports both GAAP and adjusted versions, give GAAP and note the adjusted figure separately.

The quotes step is Anthropic's quote-first technique: find the evidence before answering, so the answer is built from it. Asking for the unit heading alongside each quote is the specific fix for the FinanceBench failure, where the right page still produced wrong answers. The GAAP line exists because adjusted figures are often more prominent in a report than the statutory ones, and a model reaching for the most prominent number will take the adjusted one.

Build a bridge explaining the change in [gross margin] from [period 1] to [period 2], using the attached management accounts.

Start from the reported figure for each period [REPORTED]. Break the change into drivers available in the data (volume, price, mix, cost per unit, one-offs). Calculate each driver with code and show the formula. The drivers must sum exactly to the total change; show the residual if they don't, and don't hide it in "other".

For each driver, add one line of commentary. If the explanation isn't in the documents, write it as [ASSUMPTION] and say what evidence would confirm it.

The "must sum exactly" rule is the cheapest accuracy check in finance: a bridge that doesn't reconcile has an error in it, and asking for the residual makes the model show it instead of smoothing it away. The commentary rule matters because a margin story is exactly the kind of plausible narrative a model will write with or without evidence. Tagged, at least you can see which parts are story.

These metrics cause trouble because the names are standard and the definitions aren't.

Calculate [customer lifetime value / CAC payback / ROAS] for [segment, period] from the attached data.

Before calculating:
- State the formula you'll use and list the alternatives that would change the result.
- Say which of revenue, gross margin or contribution margin you're using, and why.
- For anything with a time horizon or discount rate, state it.
- Mark any input not in the data (e.g. a margin %, a retention rate) as [ASSUMPTION] with the value you used.

Then calculate with code and give a sensitivity table showing the result if each assumption is 20% higher or lower.

Why it matters, metric by metric:

• Customer lifetime value. Gupta and Lehmann's formulation uses margin, retention and a discount rate (MIT Sloan Management Review). Using revenue instead of margin "can inflate CLV to several times its real value" (Wikipedia, citing Farris et al.). GA4's lifetime figures are revenue, undiscounted.
• ROAS. Google Ads defines it as conversion value divided by spend, expressed as a percentage; ROI is profit divided by spend (Google Ads Help). A ROAS of 4 means four times spend in revenue. At a 25% gross margin, that's break-even before any other costs.
• CAC payback. Whether the payback uses revenue or gross margin can double or halve the months.

The sensitivity table is there because unit economics are mostly assumptions. Seeing that a CLV figure falls by a third when retention dips is worth more than the figure.

Build a [12-month] forecast of [revenue / cash] from the attached actuals.

Start with an assumptions register: a table of every assumption (growth rate, churn, pricing, hiring, payment terms), its value, its source (a figure in the data, a stated plan, or your estimate) and how confident you are. Estimates must be labelled as estimates.

Then build the forecast in code, referencing only the register. Show base, downside and upside cases, changing only the assumptions you name.

A forecast written as prose buries its assumptions in sentences. A register makes them a list you can argue with, and it's the part your CFO will ask about anyway.

Benchmarks put numbers on the caution. On Vals AI's Finance Agent benchmark (v1.1, 537 entry-level analyst tasks), the top model, Claude Opus 4.7, scored 64.37% (Vals AI, updated 4 June 2026). The best system gets roughly one in three analyst tasks wrong. Vals has reportedly released a v2 with scores that aren't comparable, so check which version you're reading. In the Datarails survey, finance teams reported spending 26% of their working week checking AI output.

Regulators and standard-setters haven't moved the responsibility. FINRA says its rules "apply when member firms use AI" and expects supervision to cover "reliability and accuracy" (Regulatory Notice 24-09). Introducing the AICPA's new practice aid, its chief auditor was plain that the auditor "remains responsible for the audit and the audit opinion" (Journal of Accountancy, October 2026).

The two-minute tie-out

1. Pick the biggest number in the answer and find it in the source: same line item, same period, same unit.
2. Recompute one derived figure (a margin, a growth rate) from its [REPORTED] inputs.
3. Check a total reconciles. Bridges sum, segments add to group, balance sheets balance.
4. Read every [ASSUMPTION] and ask whether you'd defend it in the meeting.
5. Ask: "Which figures did you not find in the documents?" Anything it now admits to should already have said NOT IN SOURCE.

Doing that for every claim in a board pack is the slow part. Count connects to Claude or ChatGPT and labels each claim as a citation to a source you can open, an assumption, model knowledge or a hallucination, with the reasoning chain from document to number. For choosing a model, see the most accurate AI for finance.

Can ChatGPT do financial analysis?

What are the best ChatGPT prompts for finance?

Which AI is best for financial analysis?

How do I stop AI inventing numbers in a financial model?