Financial and Market Research
Analyze SEC Filings With AI: A Source-Cited Workflow for Analysts
Analyze 10-Ks, 10-Qs, 8-Ks, and exhibits with AI without losing the source trail. Follow a citation-first workflow for extracting figures, comparing periods, testing claims, and producing an analyst-ready brief.
To analyze SEC filings with AI, use the model for locating, extracting, comparing, and questioning filing content, but verify every material conclusion against the original SEC filing. The minimum evidence trail is: company, filing type, reporting period, filing date, SEC link or accession number, section, page or table, and exhibit reference where relevant.
A good workflow has five linked stages: assemble the filing set, establish company and period context, extract evidence, compare changes, then draft a cited analyst brief. AI can compress the document-review cycle; it does not replace accounting judgment, valuation work, or verification.
The practical shift is to treat AI as a research layer, not the research record. A permanent workspace such as Otio for AI financial analysis is useful after SEC or web discovery because filings, earnings-call transcripts, investor decks, spreadsheets, notes, and model outputs can stay together instead of disappearing into a one-off chat.
Updated November 13, 2026.
How to analyze SEC filings with AI without losing the evidence trail
The basic mistake is asking an AI system, “What matters in this 10-K?” before defining what evidence would count as an answer. SEC filings are long, repetitive, and full of cross-references. A fluent summary can miss the table that matters, confuse fiscal periods, or cite the wrong version of an amended document.
Use this five-stage workflow instead:
Assemble the filing set. Gather the latest 10-K, relevant 10-Qs, 8-Ks, exhibits, prior comparable filings, and selected management commentary.
Establish context. Record the company, reporting period, filing type, filing date, fiscal calendar, segment structure, and any amended filings.
Extract evidence. Pull figures, disclosures, footnotes, management claims, and contractual details with direct references.
Compare and interpret. Align periods and units, calculate changes, locate management’s explanation, and test that explanation against the statements and footnotes.
Draft the brief. Write the answer around the analyst question, with citations for every material claim and a source list that can be audited.
The SEC’s own investor bulletin describes the Form 10-K as offering “a detailed picture of a company’s business, the risks it faces, and the operating and financial results for the fiscal year,” which is exactly why it should be treated as a source document rather than a summary target alone: SEC Investor Bulletin: How to Read a 10-K.
AI helps most when the question is narrower than “analyze the company.” For example:
Why did gross margin change?
Has liquidity risk increased?
Did debt terms or covenants change?
Which risk factors are new or materially revised?
Does management’s explanation match cash-flow data?
Did segment performance improve, or did one segment mask another?
Are non-GAAP claims supported by GAAP results and reconciliations?
The durable rule: every conclusion should trace back to a filing location. “Revenue increased because demand improved” is not analysis unless the source trail shows the revenue figure, the period comparison, and the disclosure that supports the demand explanation.
Choose the right SEC filings and define the analyst question first
Different filings answer different questions. AI works better when each source has a job.
A useful filing set usually starts with:
10-K: annual business overview, risk factors, MD&A, audited financial statements, accounting policies, footnotes, and long-term context.
10-Q: interim financial statements, quarterly MD&A, recent liquidity or working-capital changes, and updates since the annual filing.
8-K: material events, acquisitions, executive changes, financing events, earnings releases, impairments, restructurings, or other current disclosures.
Exhibits: credit agreements, merger agreements, employment agreements, supply contracts, amendments, and detailed supporting documents.
Proxy materials: board composition, governance, executive compensation, shareholder votes, and related-party information.
Earnings-call transcripts and investor presentations: management commentary and framing, which can be compared with the filed record but should not be treated as the same evidentiary class.
Corporate Finance Institute gives a plain overview of SEC filings as formal documents submitted by public companies and other market participants, including periodic reports and financial statements: CFI SEC filings overview. For 10-K work specifically, the SEC bulletin is the better anchor because it explains the filing’s structure and investor use.
Before asking AI to synthesize anything, create a source register. Keep it boring and complete:
Company name and ticker
Filing type
Filing date
Reporting period
SEC link or accession number
Document version, including whether it is amended
Relevant sections
Relevant exhibits
Notes on non-filed materials, such as investor decks or transcripts
This prevents a common failure: mixing a filed 10-Q number, a non-GAAP investor presentation number, and an earnings-call explanation as if they came from one source. They may all matter, but they do not carry the same status.
A practical starting corpus for one company is:
Latest 10-K
Latest two or three 10-Qs
Prior-year comparable 10-K or 10-Q
Relevant 8-Ks since the last annual filing
Material exhibits tied to debt, acquisitions, litigation, restructuring, or major contracts
Latest earnings-call transcript and investor deck if the question requires management commentary
For broader investment or diligence work, this filing set often sits inside a larger evidence base. The same source-discipline applies to market reports, customer data, expert calls, and competitor filings; the mechanics are similar to building a cited due diligence evidence base, except SEC filings give you a more formal source spine.
Prepare filings for reliable AI extraction
SEC filings are not clean essays. They contain repeated risk language, nested footnotes, dense tables, cross-references, amended versions, signatures, exhibits, and sometimes pages that extract poorly. If the text layer breaks, the AI answer may still sound confident.
Preparation is not administrative overhead. It is error reduction.
Preserve the original PDF or HTML filing. If you create a searchable copy, do not strip page numbers, section headings, table labels, footnote titles, or exhibit identifiers. Those location markers are how the work becomes auditable.
For each extraction task, ask the AI to do four things before summarizing:
Identify the relevant filing section.
Quote the surrounding passage or table label.
Report the value and period.
State whether the value is reported, calculated, or inferred.
That last distinction matters. “Revenue was $X” is a reported fact if it appears in the filing. “Revenue grew Y%” may be a calculation from two reported values. “Growth was caused by higher pricing” is an interpretation unless the filing supports it.
The extraction structure should stay consistent:
Metric or disclosure: what you are looking for.
Value and period: the exact number, date, quarter, fiscal year, or event period.
Source document: filing type and filing date.
Location: page, item, section, table, footnote, or exhibit.
Direct quotation: the surrounding source language.
Status: reported, calculated, or inferred.
Ambiguity: anything unresolved, such as units, restatements, changed segments, or amended language.
This is where a workspace matters more than a single chat window. Otio’s AI PDF reader can keep SEC PDFs in the same library as web links, spreadsheets, notes, transcripts, and decks, with reader tools for search, highlighting, and asking questions against selected text. The point is not that the AI answer is self-validating. The point is that the source and the answer remain close enough to check.
If the filing includes scanned material or a weak text layer, use OCR carefully and spot-check the extracted table against the original. Parentheses, negative signs, totals, and column headers are easy to corrupt. A shifted column can turn a useful answer into a dangerous one.
[[OTIO_INLINE_PROMO:%7B%22title%22%3A%22Ready%20to%20build%20a%20cited%20filing%20evidence%20matrix%3F%22%2C%22description%22%3A%22Keep%20original%20PDFs%2C%20exhibits%2C%20notes%2C%20and%20extracted%20claims%20together%2C%20then%20check%20each%20figure%20against%20its%20filing%20section%2C%20table%2C%20or%20footnote.%22%7D]]
Extract financial figures, disclosures, and management claims with citations
The extraction stage should separate three categories that AI often blends together:
Reported facts: figures or statements directly present in the filing.
Derived calculations: growth rates, margins, ratios, changes, or reconciliations calculated from reported facts.
Interpretations: explanations, risk assessments, or causal claims that require supporting language.
An evidence matrix is the simplest way to keep this straight. It does not need to be fancy. Use a spreadsheet, database, or structured note with fields for:
Analyst question
Claim
Value
Period
Source document
Section, page, table, footnote, or exhibit
Direct quotation
Calculation
Confidence
Unresolved ambiguity
For a gross-margin question, one row might contain reported revenue. Another contains reported cost of revenue. A third contains the calculated margin. A fourth contains management’s explanation from MD&A. A fifth contains a footnote or segment disclosure that supports, complicates, or contradicts the explanation.
The SEC’s Financial Reporting Manual describes MD&A as providing “a narrative explanation of a company’s financial statements” that enables investors to see the company through management’s eyes: SEC Financial Reporting Manual, Topic 9. That language is a useful discipline: MD&A is management’s explanation, not automatic proof of causation.
Cover the areas that materially affect the question. Depending on the company and sector, that may include:
Income statement: revenue, gross margin, operating expenses, restructuring, impairment, interest expense, tax rate.
Balance sheet: cash, receivables, inventory, debt, lease liabilities, deferred revenue, goodwill, working capital.
Cash flow: operating cash flow, capital expenditures, free cash flow calculations, stock-based compensation add-backs, changes in working capital.
Segments: segment revenue, segment profit, geographic exposure, product-line changes.
Debt and liquidity: maturities, interest rates, covenants, revolver capacity, going-concern language.
Leases and commitments: future obligations, variable payments, guarantees.
Equity and share count: dilution, repurchases, options, RSUs, convertible instruments.
Risks and contingencies: litigation, customer concentration, regulatory exposure, supply-chain risk, cybersecurity, material weaknesses.
Subsequent events: developments after the reporting period but before filing.
The AI should label GAAP figures, non-GAAP measures, management-defined metrics, estimates, and investor-presentation figures separately. Non-GAAP metrics can be useful, but the evidentiary question is different: what was reported under GAAP, how did management adjust it, and where is the reconciliation?
Use targeted follow-ups. “Analyze liquidity risk” is too broad for extraction. Better questions look like:
Identify debt maturities due within 12, 24, and 36 months, with source locations.
Extract revolver capacity, outstanding borrowings, and covenant language from the latest filing and exhibits.
Compare operating cash flow and capital expenditures for the latest period and the prior comparable period.
Quote any going-concern, covenant compliance, or liquidity-risk disclosure.
List any new or materially changed liquidity-related risk factors.
The major failure modes are predictable:
Tables extract with shifted columns.
Parentheses around negative numbers disappear.
Units change from thousands to millions.
A nine-month figure is compared with a three-month figure.
Fiscal 2026 is compared with calendar 2026 without noticing the fiscal year-end.
The model cites MD&A while using a number from a financial statement footnote.
The answer cites the correct filing but the wrong period.
An amended filing is ignored.
A good extraction workflow assumes those errors will happen and makes them cheap to catch.
Compare periods and filings without confusing correlation with explanation
Period comparison is where AI can help most and mislead fastest. The correct sequence is:
Align the periods.
Align the units.
Calculate the change.
Locate management’s explanation.
Test the explanation against statements, footnotes, exhibits, and subsequent filings.
Record what remains unexplained.
Do not start with the narrative. Start with the measurement.
For example, if gross margin improved, first confirm revenue, cost of revenue, gross profit, units, and period. Then calculate the margin and the change. Only then ask what the filing says about pricing, mix, volume, input costs, restructuring, utilization, warranty costs, inventory charges, or accounting changes.
A change log is useful for every filing comparison. Ask AI to produce a row for each material movement, but require a citation for each row. Include:
Metric or disclosure
Current-period value
Prior-period value
Absolute change
Percentage change, if meaningful
Management explanation
Supporting or conflicting footnotes
Analyst note
Verification status
This is also where multiple documents matter. A favorable headline metric can coexist with weaker economics. Revenue may grow while cash conversion deteriorates. Adjusted EBITDA may improve while debt rises. EPS may benefit from buybacks while operating income stalls. Segment growth may hide customer concentration. A lower expense ratio may reflect capitalization, restructuring, or a changed estimate rather than operating leverage.
AI can surface those tensions if asked explicitly. It will often smooth them away if asked for an “investment conclusion.”
Compare the latest filing with the prior comparable filing, not just the immediately prior filing. Quarterly sequencing matters, but seasonality can make quarter-to-quarter comparisons misleading. For some questions, trace the same disclosure across several filings to separate one-period noise from a recurring pattern.
Management commentary belongs in the comparison, but as one source. A useful next step is to compare filed disclosures with spoken explanations from earnings calls. Otio’s earnings-call transcript analysis workflow is a good companion process for checking whether the narrative on the call matches the 10-Q, 10-K, 8-K, and exhibits. Do not merge the two source types without labels: a call transcript is not the filed financial statement.
If the work includes many filings or exhibits, the process starts to resemble multi-PDF extraction. The mechanics of retaining page references, source labels, and direct quotations are close to the workflow for extracting data from multiple PDFs with source references.
Validate AI findings and produce an analyst-ready SEC filing brief
The final pass is claim-level verification. Open the cited filing. Confirm the number. Confirm the period. Inspect the surrounding table, footnote, or exhibit. Recalculate any derived metric. Mark unsupported or ambiguous statements before drafting.
Use a second model as a critic, not as a judge. A second model can search for omitted caveats, contradictory disclosures, unit errors, overconfident causal language, or missing exhibits. Model agreement is not independent verification; it is another way to generate checks.
Otio’s multiple AI models feature is useful for this kind of second-pass critique because the same project can support different model roles: one model extracts, another challenges, and the analyst verifies against the source. The source still wins.
An analyst-ready SEC filing brief should be structured around the question, not around the filing’s table of contents. A practical format is:
Executive answer: the short answer to the analyst question, with caveats.
Key evidence: the three to seven claims that matter most, each cited.
Period comparison: current versus prior comparable period, with calculations.
Management explanation: what management says and where it appears.
Cross-checks: footnotes, exhibits, cash-flow data, segment data, or later filings that support or complicate the explanation.
Risks and caveats: material uncertainties, accounting issues, one-time items, concentration, liquidity, covenants, contingencies.
Unresolved questions: items needing model work, accounting review, management follow-up, or further filings.
Source list: filing type, filing date, reporting period, SEC link or accession number, and relevant locations.
Calculation appendix: formulas, inputs, and any assumptions.
Keep the audit trail in the same project space: source files, prompts, extracted tables, calculations, notes, and final revisions. This is the difference between “AI helped me read faster” and “I can defend this analysis six weeks later.”
For teams evaluating whether a dedicated workspace fits this process, check Otio pricing against the size of the source sets you typically handle: number of filings per project, PDFs and exhibits, spreadsheets, transcripts, and whether second-pass model critique is part of the workflow.
The best next step is small. Pick one company and one question. Assemble the latest 10-K, the latest comparable 10-Q, the prior-year comparable filing, and any relevant 8-K or exhibit. Build an evidence matrix for three material claims. Verify every citation manually. Then expand the corpus.
FAQ
Q: Can AI accurately analyze SEC filings?
A: AI can help extract, compare, and organize information from SEC filings, but its output must be checked against the original filing. Tables, footnotes, reporting periods, units, and amended documents are common sources of error.
Q: Which SEC filings should analysts analyze together?
A: A useful starting set is the latest 10-K, the latest two or three 10-Qs, relevant 8-Ks and exhibits, and the prior-year comparable filing. Add earnings-call transcripts or investor presentations when comparing filed disclosures with management commentary.
Q: How do you cite AI-generated analysis of an SEC filing?
A: Cite the original filing rather than the AI response. Record the company, filing type, filing date, reporting period, SEC link or accession number, and the relevant page, section, table, footnote, or exhibit.
Q: Can AI replace financial analysis or investment judgment?
A: No. AI can accelerate document review and surface relationships, but analysts still need to validate calculations, assess accounting context, evaluate risks, and make or communicate judgments responsibly.
[[OTIO_FOOTER_PROMO:%7B%22title%22%3A%22Apply%20this%20workflow%20to%20your%20filing%20set%22%2C%22description%22%3A%22Bring%20your%20company's%2010-Ks%2C%2010-Qs%2C%208-Ks%2C%20exhibits%2C%20notes%2C%20and%20transcripts%20into%20one%20library%2C%20then%20draft%20a%20brief%20with%20citations%20for%20every%20material%20claim.%22%7D]]
