Use AI for financial research by giving it a bounded question, an explicit source set and a record of what each source supports. Verify the important claims and calculations yourself before acting on the result. The useful output is an evidence-backed research note with visible gaps, not an answer whose fluent tone substitutes for checking.
Suppose you want to understand a payment company’s revenue. A model can help organise a filing and identify terms to investigate. It can also confuse gross payment volume with revenue, merge figures from different periods or supply a citation that does not support the sentence beside it.
This guide sets out a reusable workflow. The examples and prompts are illustrative; they are not a benchmark of any particular model or a recommendation to invest in a company.
Begin with a question that can be answered
“Is this a good investment?” is too broad for a first research task. It invites a model to blend facts, assumptions and an implied recommendation. Break the problem into questions whose evidence you can inspect.
For example: What does the company call revenue? Which business segments contributed to it? How did operating cash flow change between two specified years? What risks does management describe, and which claims come from outside management?
Write the entity, period, currency and intended use at the top of the research note. A company group, a subsidiary and a token issued by an associated organisation are not interchangeable subjects. The scope should prevent the model from quietly borrowing evidence from the wrong one.
State the decision boundary too. If you are trying to understand a business model, the output need not include a price target. If you are comparing transfer costs, it needs current executable quotes rather than a general description of the providers.
Collect sources before asking for a conclusion
Start with documents appropriate to the claim: official filings for reported financial statements, regulator pages for regulatory status, issuer terms for redemption rules and product documentation for a feature. A company’s own statement is evidence of what it says, not independent proof of every claim it makes.
Keep commentary and primary evidence distinguishable. A journalist’s interpretation can help identify questions, but verify consequential figures against the underlying document when available. Record the publication date and the period covered; those are often different.
For a financial filing, keep the document title, page or section and reporting unit. “Revenue 120” is unusable without knowing whether the number means dollars, thousands or millions and whether it covers a quarter or year.
The SEC’s introduction to financial statements is useful background for distinguishing income, balance-sheet and cash-flow measures. That conceptual preparation makes it easier to notice when a generated summary puts unlike figures side by side.
Give the model a narrow extraction job
Ask for a table of evidence before asking for a narrative. A useful instruction is:
Use only the supplied documents. For each requested metric, return the value, unit, period, document title and exact page or section. Mark missing values as “not found.” Keep reported values separate from calculations. Do not infer a missing number from an unrelated period.
The instruction improves the shape of the task; it does not guarantee obedience or accuracy. Open the cited locations and compare the values. If the document has tables spanning pages, check whether the model missed a column heading or carried the wrong year across a page break.
Ask the model to preserve the company’s terminology initially. Renaming a metric can hide differences between an accounting measure and a company-defined measure. You can explain the terminology later, after establishing what was actually reported.
Avoid asking for too many unrelated outputs at once. Extracting three clearly defined metrics is easier to review than receiving a long report with dozens of claims whose sources are scattered through the text.
Build a claim ledger
A claim ledger connects each conclusion to its evidence and records whether it is a fact, calculation or interpretation. It also gives you somewhere to keep uncertainty without burying it in the final prose.
| Claim type | Example | Evidence needed |
|---|---|---|
| Reported fact | Revenue was a stated amount in a stated year | Filing, table, unit and period |
| Calculation | Revenue increased by a calculated percentage | Both inputs and the formula |
| Interpretation | Growth appears concentrated in one segment | Segment evidence and limits of the inference |
| Unresolved question | Whether a cost will recur | Missing evidence or competing explanations |
Use a status column such as verified, needs checking or unsupported. “Verified” should mean someone checked the underlying evidence, not that another model agreed with the first answer.
When a source supports only part of a claim, split the sentence. A filing may establish that costs increased, while the suggested explanation for that increase remains an inference. Keeping those pieces separate makes the final note easier to challenge and update.
Recalculate every consequential number
Do arithmetic with a calculator or spreadsheet whose inputs you can inspect. A model can explain a formula, but the final result should be reproducible independently of its prose.
Take a fictional example: revenue rises from $80 million to $100 million. The increase is ($100m − $80m) ÷ $80m = 25%. Dividing by the new value instead would produce 20%, which answers a different question.
Now suppose the $100 million figure covers 15 months and the $80 million figure covers 12. The arithmetic is still correct for the raw figures, but calling it comparable annual growth would be misleading. Checking the period matters as much as checking the formula.
Keep reported and adjusted measures in separate rows. If management excludes certain costs from an adjusted result, record those exclusions. Do not combine an adjusted numerator with an unadjusted denominator unless you can explain why the comparison is meaningful.
For currency conversions, write the rate direction, date and purpose. A period-average rate and a closing-date rate are not interchangeable merely because both convert dollars to euros.
Test citations as claims, not decorations
Open each source supporting a consequential statement. Confirm that it exists, concerns the right entity and period, and supports the wording actually used. A real URL can still be a bad citation.
Watch for a source that is older than the claim, a summary citing another summary, or a document that contains the topic but not the asserted fact. These errors can survive a superficial check because the link looks plausible.
If a source is inaccessible, label the limitation and look for an accessible primary version. Do not tell the reader that you verified a document you only saw mentioned in search results.
The NIST Generative AI Profile discusses confidently incorrect output, information integrity and human reliance on generated content. The practical implication for this workflow is to make evidence inspection an explicit step, rather than hoping a confident answer is a correct one.
Ask for counterevidence and missing information
Once the basic facts are checked, ask what would weaken your current interpretation. This is particularly useful when the first summary tells a satisfying story.
For example, increasing revenue and positive profit might suggest improvement, but receivables could be growing faster than sales. That observation does not prove a problem; it creates a question about collection timing and revenue quality. Our profit and cash flow guide explains the underlying distinction.
A useful prompt is: “List the strongest alternative explanations supported by these documents. For each, identify the evidence that would help distinguish it from the current interpretation. Do not invent missing evidence.”
Keep unanswered questions in the final note. Research quality is not measured by eliminating every blank. It is measured partly by recognising which blanks matter to the decision and refusing to fill them with a guess.
Protect the material you submit
Public filings are different from confidential budgets, client invoices and unpublished transaction records. Before sending private material to an AI service, check the service arrangement, access controls, retention settings and whether you are authorised to share it.
Removing a name may not make a document anonymous. Account details, invoice references, unusual transaction amounts or combinations of contextual facts can still identify a person or business.
For workflow experiments, start with synthetic data. If a task can be demonstrated with fictional invoices, there is no need to upload a customer’s real payment instructions just to see whether a table can be extracted.
Treat retrieved documents as evidence, not instructions to the assistant. Text inside a PDF or website should not be allowed to redirect the workflow, request credentials or authorise payments. The invoice review workflow applies that separation to a concrete administrative task.
Produce a short research memo
After checking the evidence, ask for a memo organised around the original question. It should state the answer, supporting facts, calculations, uncertainties and the next evidence needed. Keep recommendations separate from observations.
A useful structure is: scope and date; findings; calculation table; alternative explanations; unresolved questions; source list. The structure serves the research rather than forcing every topic into an identical essay.
Ask the model to label inferences explicitly and preserve limitations from the ledger. Then read the final prose against the ledger. A common editing failure is turning “may reflect” into “was caused by” because the stronger sentence sounds cleaner.
Save the evidence and working calculation with the memo. If the source later changes or a new filing appears, you should be able to see which conclusion needs review instead of repeating the entire investigation.
Decide what the result is ready for
A checked educational note can help you understand a company or prepare questions for an adviser. It is not automatically a personalised investment recommendation, a legal opinion or a complete due-diligence report.
Match the review effort to the consequence. A rough topic outline needs less verification than a number used to value a business or approve a major payment. When a decision depends on specialised accounting, tax or legal interpretation, bring the evidence to a suitably qualified professional.
The useful role for AI is to reduce the mechanical work of organising documents and surfacing questions. You remain responsible for deciding whether the evidence supports the claim and whether the remaining uncertainty is acceptable for the intended use.
Questions
Can I trust an AI answer if it includes sources?
Only after checking the relevant sources and their support for the claims. A real link can concern the wrong period, entity or statement.
Should I ask a second model to verify the first?
A second model can suggest objections, but agreement between models is not verification. Check the primary evidence and reproduce the calculations.
What should I do when a metric is missing?
Mark it as missing and explain why it matters. Do not substitute a different period or company-defined measure without making the change explicit.
Can I upload private financial documents?
Check your authority to share them and the service’s data arrangement first. Use synthetic or properly minimised data when it can answer the workflow question.
