How do you fact-check an AI answer about a company?
Break the answer into single claims, then check each figure, name and date against the filing or register it should come from. The check matters: in Patronus AI's FinanceBench test of questions drawn from SEC filings, GPT-4-Turbo with retrieval answered incorrectly or refused 81% of a sample (Patronus AI).
Why do AI answers about companies go wrong?
General assistants predict likely text, so a confident figure can be out of date, from the wrong entity or made up. FinanceBench, released in 2023, used 10,231 questions about public companies; even the best setup that read nearly a whole filing got 79% of answers right (Fortune, December 2023). A 2026 benchmark by Saturn put 121 money questions to 18 models and found average accuracy of 43%, falling to 12% on hard questions (Wealth Professional, 2026).
Models have improved since, but the failure types repeat. Watch for figures from the wrong fiscal year, a parent company confused with a subsidiary of a similar name, a private company's revenue presented as fact when no filing exists, and links that point to pages which do not contain the claim.
Step 1. Split the answer into claims
Copy the answer and put each factual statement on its own line: every number, every name, every date and every "the company owns" or "the company raised". Even a short answer usually holds more checkable claims than it seems. Opinions and summaries need no check; the facts under them do.
Step 2. Pin down the exact legal entity
Most errors start here. Find the company's legal name and an identifier before you check anything else. For US public companies use the CIK number on SEC EDGAR; for any company with one, the LEI on GLEIF; for UK companies the company number on Companies House. How to find a company's LEI shows the GLEIF route. If the AI answer mixes a brand name with several entities, treat every figure as unconfirmed.
Step 3. Match each claim to the record it should come from
Each type of claim has a primary source. Check it there, not in another summary.
| Claim | Primary source | Where to look |
|---|---|---|
| Revenue, profit, debt of a US public company | Form 10-K or 10-Q on SEC EDGAR | Item 8 financial statements, Item 7 MD&A |
| A recent event, deal or executive change | Form 8-K on SEC EDGAR | The item number for that event |
| Private fundraising in the US | Form D on SEC EDGAR | Items 13 and 14, amount sold |
| Subsidiaries of a US public company | Exhibit 21 to the 10-K | The list of subsidiaries and jurisdictions |
| Corporate parent | GLEIF Level 2 data | Direct and ultimate parent on the LEI record |
| UK company owners and accounts | Companies House | People with significant control, filing history |
| Sanctions status | OFAC, UK, EU and UN lists | The official list search |
| Lawsuits in US federal court | CourtListener | Docket search by party name |
Step 4. Check the number, the unit and the period
Open the filing and find the figure. Confirm the unit (thousands or millions), the currency and the period (fiscal year, quarter or trailing 12 months). Many AI errors are correct numbers from the wrong year. If the figure is not in any filing, as with most private company revenue, the answer is an estimate and should be labelled as one; see how to estimate private company revenue.
Step 5. Open every link the AI gave
Click each cited link and search the page for the claim. Mark the claim confirmed only if the page states it. A link to a company homepage or a search results page does not support a specific figure. A secondary article that repeats a figure is weaker than the filing it came from, so follow it back to the filing.
Step 6. Record the source and the date next to each claim
Write the source link and the date you checked it next to each confirmed claim, and strike or flag the rest. Records go stale: a parent company, a director or a status can change after the filing. For a method to test freshness, see how to check that private company data is current.
Where does Tamarly Intelligence help?
Tamarly Intelligence is Tamarly's research platform on companies, funds, countries and markets, built from official filings and open data. Its Ask the Brain feature answers from Tamarly's data and knowledge graph with a numbered source on every sentence, graded A (official filing or register) to D (secondary source), so the check in step 5 starts from the record. When Tamarly does not cover a company it says so.
Get company answers with a graded source on every sentence, free for 7 days
Frequently asked questions
Is ChatGPT accurate for company financial data?
Not reliably without sources. In FinanceBench, built from SEC filing questions, GPT-4-Turbo with retrieval answered incorrectly or refused 81% of a sample (Patronus AI). Newer models do better, but every figure still needs a check against the filing.
What is the fastest way to check a public company figure?
Search the company on SEC EDGAR, open the latest 10-K or 10-Q and use your browser's find function for the figure or line item. Check the unit and period in the table header before you accept a match.
How do you check a private company figure?
Look for a primary record first: a Form D for US fundraising, Companies House accounts for UK companies, or ESEF filings for listed European groups. If no record holds the figure, treat it as an estimate and say so.
About the author. Melvine Manchau is the founder of Tamarly and author of the Tamarly Newsletter. He previously worked at PwC, Salesforce, BlackRock/eFront and Lazard.
About Tamarly Research. Tamarly Research is part of Tamarly Media, which holds no investment positions. Tamarly is a New York research house that studies how AI is changing regulated institutions.