Can ChatGPT Analyze a 10-K? A Verification Workflow
ChatGPT can read a 10-K, but the useful workflow separates extraction, calculation, and judgment while requiring a source check for every material claim.
Published August 28, 2026 · Updated August 30, 2026
In this article
Yes. ChatGPT can analyze an uploaded 10-K well enough to create a first-pass summary, extract disclosed facts, compare sections, and draft follow-up questions. The safe unit of work is a cited claim, not an uncited report. Ask the model to identify the filing section and quote a short supporting passage, then verify every number, period, and qualifier in the original filing before it enters a model or investment memo.
This article is a workflow guide based on public SEC material and current OpenAI documentation. It is not a benchmark of ChatGPT against other products, and no accuracy rate is implied.
Start with the filing, not the annual-report design
A company's glossy annual report and its Form 10-K may overlap, but they are not always the same document. Use the filing from EDGAR as the source of record. The SEC's 10-K reading guide explains the standard sequence: Business, Risk Factors, MD&A, audited financial statements, controls, exhibits, and related disclosures.
The SEC also provides EDGAR search guidance. A domestic issuer's 10-K includes audited annual financial statements, material risk factors, and management's discussion of results. Inline XBRL may expose many financial facts in structured form, which is useful when a table in the document is difficult to parse.
Before uploading anything, record:
- company name and CIK;
- filing type and accession number;
- fiscal year end and filing date;
- whether the year contains 52 or 53 weeks;
- currency and units used in each table;
- amendments or incorporated-by-reference material;
- whether the document is public and permitted under your firm's policy.
That header prevents a common category of error: a correct number attached to the wrong period, unit, or document.
A four-pass workflow
Do not ask for a complete investment view in one prompt. Separate the job into passes with different evidence standards.
| Pass | What ChatGPT should produce | Analyst check before continuing |
|---|---|---|
| 1. Map | Filing outline, fiscal period, units, and relevant sections | Match the outline to EDGAR and confirm the reporting calendar |
| 2. Extract | Small table of disclosed facts with section and page or anchor | Open each source and check sign, unit, segment, and qualifier |
| 3. Calculate | Explicit formula using verified inputs | Recalculate in a spreadsheet and inspect denominator choices |
| 4. Interpret | Questions, changes, and possible explanations | Separate management's explanation from the analyst's inference |
Pass 1: map the document
Ask the model to list the sections it can identify and to flag anything it cannot read. This is a capability check, not busywork. If the model cannot locate the financial statements, notes, or exhibits, stop and use the EDGAR HTML filing or extract the relevant pages.
For a live example, Costco's fiscal 2025 Form 10-K identifies its reporting calendar and presents comparable-sales measures with defined exclusions. Those details should be captured before any year-over-year comparison is calculated.
Pass 2: extract narrow facts
Use a table with one claim per row. A useful schema is:
| Field | Required content |
|---|---|
| Claim | One factual statement, without interpretation |
| Value | Number plus currency, unit, and sign |
| Period | Quarter, fiscal year, or point-in-time date |
| Scope | Consolidated, segment, geography, or product |
| Source | Filing item, note, table label, and page or HTML anchor |
| Qualifier | Adjustment, exclusion, accounting definition, or management caveat |
Require “not found” when the filing does not support a request. A blank is more useful than a plausible completion.
Pass 3: calculate in public
Ask for the formula and the exact source rows used. If the task is margin change, the output should expose revenue, profit, each period, and the division used. If the task is a growth rate, specify whether it is reported, organic, constant-currency, or comparable.
The model can write the calculation. The analyst should recompute it in a controlled spreadsheet or script. This keeps extraction errors separate from arithmetic errors and makes the review trail clearer.
Pass 4: interpret after the facts are locked
Create three labeled groups:
- Reported fact: what the filing or management explicitly states.
- Derived observation: a calculation from verified inputs.
- Analyst hypothesis: an explanation or investment implication that still needs evidence.
The labels keep a polished paragraph from turning an inference into a reported fact. They also make follow-up work obvious. A hypothesis about price, mix, churn, or capacity should point to the next transcript, filing note, competitor disclosure, or channel source needed to test it.
Give the model a narrow extraction brief
After attaching the filing, define the assignment in ordinary language instead of asking for a complete analysis. The following filled brief shows the level of specificity needed:
| Brief element | Instruction |
|---|---|
| Source boundary | Use only the supplied Form 10-K; do not add outside knowledge |
| Identity check | Confirm the company, form type, fiscal year end, filing date, accession number, currency, table units, and unusual week count |
| Research question | Identify the issuer-reported drivers of the year-over-year change in gross margin |
| Relevant material | Locate the financial statements, MD&A, and notes that address margin and its stated drivers |
| Output structure | Return one factual claim per row with value and unit, period, scope, filing section, page or HTML anchor, a short supporting excerpt, and any qualifier |
| Missing evidence | State “not found in the supplied filing” rather than completing the gap |
| Calculation gate | Do not infer a chart value or calculate a result until an analyst approves the extracted inputs |
After checking the extraction, make a separate calculation request that names only the approved inputs. Keep the source fields beside the result so the arithmetic can be reviewed independently from the extraction.
Where ChatGPT is useful
The product works well for navigation and first-pass synthesis when the scope is explicit. It can turn a long Risk Factors section into a categorized index, locate references to a customer or supplier, draft a table from one note, and propose diligence questions. It can also generate code for structured filing data.
OpenAI documents file inputs in the Responses API, and its file documentation explains storage and expiration controls for uploaded files. ChatGPT Enterprise adds organizational privacy, security, and admin controls, according to the current product description. Those capabilities do not decide whether a particular answer is financially correct.
Failure modes to design around
Period drift. A model may mix fiscal and calendar labels or use a comparative column from the wrong year. Put the period in every extracted row.
Qualifier loss. Comparable sales, adjusted earnings, backlog, and recurring revenue are defined by the issuer. Preserve the definition and exclusions beside the value.
Table geometry. A PDF can separate headers from cells, repeat column names, or place footnotes far from the value. Switch to EDGAR HTML or structured facts when the parsed table is ambiguous.
Cross-document leakage. A prior filing or web result may look like the supplied source. Tell the model to use only the attached filing during extraction and verify the accession number in every cited link.
False completeness. A concise answer may omit a note that changes the interpretation. Ask which relevant sections were not read and which requested fields were absent.
Narrative overreach. Management's stated driver is evidence of management's view. It is not proof of causality. Keep explanations attributed and place independent analysis in a separate field.
When one uploaded filing is not enough
A single-document workflow becomes cumbersome when the job spans a coverage list, several years, entitled broker research, internal models, or scheduled monitoring. The problem then shifts from model capability to data management, permissions, and repeatability.
A research platform can pre-index filings and connect each answer to a stored company, period, source, and workflow. See what an AI investment research platform is for the system boundary. AllMind, our own product, is one vendor in that category, and we wrote this guide. Our platform page describes document search and grids that apply one question across a universe, over SEC and EDGAR filings, SEDAR+, and global filings from the 18 markets we cover. Those are our first-party claims; this guide does not present a product run.
Stay with ChatGPT when the assignment is occasional, the filing is public, and the analyst can check every output directly. Evaluate a team platform when the same extraction must run across many names, retain internal context, respect content entitlements, or leave a durable audit record.
The verification checklist
Before copying any output into research of record, confirm:
- The accession number and fiscal period match the assignment.
- Every material number has a unit, period, scope, and openable source.
- Table headers and footnotes were checked in the original filing.
- Non-GAAP and company-defined metrics retain their definitions.
- Derived figures were recomputed outside the chat.
- Reported facts, calculations, and analyst hypotheses are labeled separately.
- Missing evidence appears as “not found,” without a guessed completion.
- The uploaded document and output comply with the firm's data policy.
- A named analyst owns the final judgment.
The practical answer is therefore conditional. ChatGPT can do useful 10-K analysis when the analyst narrows the task and verifies at claim level. It should produce a research draft with a visible source trail, never an unreviewed investment conclusion.
Sources and methodology
This workflow was revised on August 30, 2026 using the SEC's 10-K guide, EDGAR research guidance, the public Costco fiscal 2025 filing, and current OpenAI file documentation. We did not execute or measure a controlled ChatGPT run for this revision. Product behavior can vary by model, plan, file format, and account settings.