An Evidence-First Workflow for AI Financial Research
In this essay
The easiest way to get a compelling AI investment memo is to ask for one. The harder task is discovering whether the memo’s claims survive contact with the underlying documents.
A better workflow begins with a question small enough to investigate. Instead of “Is this a good AI stock?”, try: “Did growth in reported revenue turn into operating cash, and what spending was required to support it?” The narrower question makes evidence requirements visible before a narrative takes over.
1. Write the question and its boundary
Specify the business, periods, metrics, and decision the research will inform. Write down what the analysis will not be able to establish. A company may discuss AI enthusiastically while reporting no separate AI revenue line. The absence of a disclosure is a finding, not permission to fabricate one.
For the cash-conversion question, a basic scope might cover three reported fiscal years, operating cash flow, capital expenditure, and the relevant working-capital notes. A different business may require a different set of measures. Financial institutions, for example, do not fit every industrial-company cash-flow shortcut.
The scope is a research plan. It should be adjusted when the documents reveal a better question, with the change recorded rather than quietly forgotten.
2. Build a source ledger before a summary
Start with the issuer’s actual filings and disclosures. For US public-company research, the SEC’s guide to reading a 10-K/10-Q explains where to find business descriptions, management discussion, financial statements, and notes.
Record the document title, filing or release date, reporting period, URL, and section used. If a filing is amended, retain the version information. A presentation can help explain management’s framing, but a promotional slide is not interchangeable with a financial statement.
The SEC’s guide to financial statements is a useful starting reference for distinguishing income, balance-sheet positions, and cash flows. In a research ledger, keep those distinctions in the metric names rather than relying on memory.
3. Extract claims into a small evidence table
Ask the assistant to propose candidate facts with exact locations. Require an explicit “not disclosed” when the source does not answer the question. Then inspect the material facts yourself.
| Field | What to preserve |
|---|---|
| Fact | What the document actually states |
| Scope | Entity, period, unit, and accounting basis |
| Evidence | Document and table or section |
| Calculation | Formula using checked inputs |
| Judgment | Interpretation and competing explanations |
This table is a proposed working format. Its value is that a reader can challenge a judgment without losing track of the facts underneath it.
Suppose management says customers are experimenting with a product. Rewriting that as “adoption is accelerating” adds a trend and a commercial conclusion. Keep the original statement’s strength until additional evidence supports the stronger wording.
4. Recompute the important numbers
Use an invented two-year example, in millions:
| Metric | Year 1 | Year 2 |
|---|---|---|
| Revenue | 1,000 | 1,200 |
| Operating cash flow | 180 | 210 |
| Capital expenditure | 80 | 160 |
| Operating cash flow minus capital expenditure | 100 | 50 |
Revenue grew 20% and operating cash flow grew about 16.7%, while the simple cash-flow-minus-capex measure halved. That is not proof of deteriorating economics. The additional spending could support future capacity, replace assets, or reflect timing. The table creates a question about the spending rather than answering it.
The measure is deliberately labeled by its arithmetic. Calling it “owner earnings” would require further decisions about maintenance spending, financing, dilution, and the business model. A convenient subtraction should not quietly acquire a stronger definition.
Use a separate calculation step for material ratios. Preserve unrounded inputs, round at presentation, and check units before comparing periods.
5. Give the counterargument equal access to the documents
A general request to “be skeptical” often produces generic risks. Ask a concrete contrary question instead: “Which disclosure would be most inconsistent with the claim that growth is becoming more capital-efficient?” Then look for the actual evidence.
If the thesis depends on pricing power, inspect retention and contract commentary where disclosed. If it depends on capital efficiency, inspect spending commitments, utilization disclosures, and depreciation assumptions where relevant. Do not imply these facts exist merely because they would be useful.
Record the strongest alternative explanation and the evidence that would distinguish it. The model’s role is to generate search directions; the documents must do the work of supporting them.
6. Finish with an uncertainty register
End the memo with three compact fields: what is established, what is inferred, and what is unknown. Add a concrete trigger for revisiting the conclusion, such as a new filing, a changed metric definition, or a missing disclosure becoming available.
Avoid a numeric confidence score unless its meaning has been defined and calibrated. “High confidence” generated as prose is not a measured probability. For many research notes, a precise account of the unresolved assumptions is more useful.
Save the question, source ledger, calculations, and final memo together. A second reader should be able to reconstruct the important claims without recreating the conversation with the model.
AI is most useful here as a reading and drafting assistant with an inspectable trail. The finished product is a better-supported question or conclusion, not merely a longer document.
Further reading
The method connects to retrieval and verification, valuation assumptions, and the hidden costs of an AI workflow. All numerical figures above are fictional educational examples, not company disclosures.