A useful research assistant should make it easier to discover why an investment idea is wrong. If it merely turns a prompt into a persuasive essay, it may increase confidence without increasing knowledge.
AI is most helpful when the work can be divided into small, checkable tasks: finding a disclosure, comparing two tables, reconstructing a calculation, and identifying what evidence would change the conclusion.
A question with identifiable evidence
“Is this company attractive?” invites a broad story. “Did revenue grow because existing customers bought more, prices rose, or an acquisition added sales?” points toward specific evidence.
Give the model a bounded source set: the relevant filings, earnings releases, and transcript, with publication dates. Ask it to return the passage or table supporting each claim. A citation should point to the evidence, not merely to a homepage that happens to concern the company.
Keep three categories visible in your notes: reported facts, your interpretation of those facts, and assumptions about the future. A model can help with each, but moving a sentence from the third category into the first is not research.
From reported figures to a calculation
Suppose revenue rises from $80 million to $100 million. Growth is 25%. If operating margin rises from 20% to 22%, the improvement is two percentage points, or 10% relative to the original margin. Operating profit rises from $16 million to $22 million, a 37.5% increase.
These are three different growth measures. Confusing them can make a modest operating change sound much larger or smaller than it is. Ask for the inputs and formula, then verify the calculation in a spreadsheet or code.
The FinQA research dataset was built around questions requiring numerical reasoning over financial reports and explicit reasoning programs. Its enduring lesson for a workflow is that retrieving the right sentence and performing the right calculation are separate tasks. The paper's historical benchmark results should not be treated as a rating of current models.
What changed besides the headline number
Correct arithmetic can still support a bad inference. If $15 million of the new $100 million revenue came from an acquisition, comparing the total with $80 million does not establish 25% growth in the original business. The remaining $85 million implies 6.25% growth, assuming the periods and accounting definitions are otherwise comparable.
Likewise, operating profit is not cash available to shareholders. Capital expenditure, working-capital changes, interest, and taxes can matter. Ask the assistant to build a bridge between the reported metric and the economic quantity your thesis actually depends on.
This is a better use of a model than requesting a longer summary. Each bridge forces an assumption into the open.
Alternative explanations
Once a thesis exists, ask what observation would weaken it. If the argument is that higher margins reflect pricing power, possible alternatives include lower input costs, product mix, or temporarily reduced spending. Find the disclosures that could distinguish them.
Build a small scenario table with a few meaningful drivers rather than dozens of precise-looking forecasts. For each driver, state the source, the assumption, and the result if it is wrong. The model can generate candidate explanations; the analyst must decide which are supported.
Knowledge of later events
A model used today may know events that happened after the historical date being simulated. Asking it to “pretend it is 2020” does not prove that knowledge has been removed. Recent research on look-ahead contamination in financial AI explores this problem.
For a historical experiment, record what documents and model version were available, and distinguish a retrospective explanation from a prospective forecast. A cleaner test starts a dated prediction record now and scores it later without rewriting the original reasoning.
From research to a proposed order
A useful automated workflow can prepare an order proposal with the instrument, direction, size, price limit, and supporting evidence. Execution needs additional controls: maximum exposure, allowed instruments, stale-data checks, and a reliable way to cancel or stop.
In the revenue example, the useful result is the distinction between 25% total growth and 6.25% growth excluding the acquisition. That changes the valuation question. A confident summary that misses the distinction has done less research, however well it reads.