How to Use AI Tools for Investment Analysis: Financial Statements and Risk Models
An afternoon, same-input test can show how each assistant handles financial-statement extraction and risk-model tasks, but it cannot establish a universal winner.
Sources checked 2 Oct 2026
Use AI tools to extract financial-statement data, calculate risk metrics, and test model logic, but do not treat their output as an investment decision. The vendors’ official pages document plan, usage, data-setting, and policy terms; they do not establish accuracy on a specific investment workflow, and Chat Picker has not tested the assistants for this use. Run the same inputs and review rules across ChatGPT, Claude, and Gemini, then base your choice on your own log.
Know these limits before you start
Plan pages are not performance evidence. As read on OpenAI’s pricing page and Anthropic’s pricing page in October 2026, usage is described in relative terms, without exact message counts. Google’s US AI plans page says Gemini’s compute-based limits refresh every five hours, up to a weekly limit, and that paid tiers offer multiples of free usage. Recheck your account before uploading a statement.
For chat data, OpenAI lists an opt-out on Free, Go, Plus, and Pro. Anthropic lists an opt-out on Free, Pro, and Max and says Team is not trained on by default. Google’s plans page does not state a corresponding setting, so the comparison matrix marks it “Not verified.” Treat missing information as unresolved, not permission, and use only files you are authorized to submit.
The policies also make human review unavoidable in some circumstances. OpenAI’s Usage policies prohibit automated high-stakes decisions in sensitive areas without human review. Anthropic’s Usage Policy requires a qualified professional to review covered advice, recommendations, and subjective decisions before finalization. Gemini’s safety and policy guidelines say factually inaccurate outputs that could cause significant financial harm should not be generated and note that language models are probabilistic. Keep professional and compliance review in the workflow.
The test plan
Use an authorized, fixed document packet. If an assistant rejects a file or format, record that outcome rather than changing the input for only that product.
- Extract statement values. Prepare a spreadsheet containing a balance sheet, income statement, and answer key with the expected period, unit, sign, and page or cell for each tested line item.
Prompt: “Use only the attached financial statements. Return one row per line item with statement, section, period, value, unit, sign, source location, and status. Preserve source units and signs. Write NOT FOUND for anything you cannot locate; do not infer it.”
Look for: Wrong periods, signs, units, unsupported values, and missing references.
Log: Each mismatch, its correct value, and the source location.
- Check risk-adjusted calculations. Prepare a dated price or total-return series, a stated risk-free rate, a benchmark series, and written metric definitions.
Prompt: “Using only the attached return series and assumptions, calculate annualized return, annualized volatility, Sharpe ratio, and maximum drawdown. Show every input, formula, substituted expression, result, and assumption. Flag missing or inconsistent data and do not infer replacements.”
Look for: Reproducible formulas, consistent annualization, correct signs, and disclosed assumptions.
Log: Formula errors, unit mismatches, unresolved inputs, and differences from your independent calculation.
- Test a multi-factor framework. Prepare a table containing leverage, liquidity, earnings volatility, market correlation, scenario labels, and predefined normalization and weighting rules.
Prompt: “Build a comparative risk table from the attached dataset. For each scenario, show every factor, input cell, normalization method, weight, contribution, and result. Do not invent weights, correlations, or missing values. State each limitation.”
Look for: Consistent definitions, duplicated factors, unsupported assumptions, and confused correlation with causation.
Log: Factor-definition issues, double counting, missing inputs, and scenario inconsistencies.
- Compare structured output. Prepare the same statements and define the exact CSV columns you need.
Prompt: “Repeat the extraction as CSV with exactly these columns: issuer, statement, section, period, value, unit, source_location, status. Quote text fields, preserve numeric signs, use FOUND or NOT FOUND in status, and emit data rows only.”
Look for: Valid CSV, stable column names, correct data types, preserved signs, and usable source references.
Log: Parse errors, schema changes, misplaced values, and formatting that requires manual repair.
- Check version sensitivity. Freeze the input packet and record the plan, visible model label, tool settings, and run time. Repeat the same task whenever the visible model or settings change.
Prompt: “Use only the attached statement set and return series. Extract every line item and calculate the supplied risk metrics. Cite each source location and label every assumption.”
Look for: Changes in extraction, formulas, citations, refusals, or formatting that coincide with a recorded change.
Log: Separate entries for each run; do not overwrite the earlier result.
- Measure your own cost and speed. Run the same extraction and calculation sequence under each plan you already have access to.
Prompt: “Use only the attached materials, complete both requested tasks, cite source locations, and identify missing data. Do not make an investment recommendation.”
Look for: Upload or usage blocks, time to accepted completion, displayed plan changes, and interruptions.
Log: Elapsed time, usage-counter changes, displayed charges, and format switches. Treat speed as workflow data, not evidence of accuracy.
How to read your results
Keep one row per run:
| Assistant | Plan | Visible model label | Format | Extraction issues | Formula issues | Source gaps | Elapsed time | Usage or charge change |
|---|---|---|---|---|---|---|---|---|
Verify every extracted value against the source document and recalculate each financial formula independently. Keep “wrong,” “not reported,” and “source unavailable” as separate categories; fluent prose should not conceal any of them.
Define your must-have conditions before choosing: accepted input, material values verified, formulas reproducible, sources traceable, and required format usable. Compare time and cost only among runs that meet those conditions. Do not combine unlike failures into an arbitrary score or treat this exercise as a general ranking.
Where the plans differ
In the Claude vs ChatGPT vs Gemini matrix, inspect the rows for Free plan, Low-cost tier, Main paid plan, Heavy-use plans, Context window in the app, How usage limits are described, and Training on your chats. Add Team plan if this is a shared workflow.
A free plan does not imply equivalent access. OpenAI says free uploads are limited; Anthropic lists web search, file creation, code execution, and memory; Google describes a free Gemini plan with an account and storage. In the same October 2026 reading, OpenAI lists context windows by model class and plan, Anthropic says context can reach 1M and varies by model, and Google’s plans page gives no app token figure.
Use the price rows to check whether added capacity is relevant to your repeated workload, then confirm the current vendor price before subscribing. The usage and training rows determine feasibility and data handling. Neither row validates investment-analysis quality, so your same-input log remains the deciding evidence.