How to Analyze a 10-K with AI

To analyze a 10-K with AI, have the tool read the entire filing (statements, footnotes and exhibits), ask it specific questions section by section, compare the filing with last year’s, and check every number it gives you against the line it came from. Kepler is built for exactly that: extraction models we train read every page of every 10-K before you ask, statements assemble from the filing’s own XBRL with no model in the path, and every figure opens to its page and line, as filed or as restated. We build Kepler. This guide covers what’s in a 10-K, the questions worth asking of each part, how to read a filing against the prior year, a 12-question checklist, prompts that work, and where AI tends to miss.

What’s in a 10-K, and what to ask of each part

A 10-K is a company’s annual report to the SEC. Large accelerated filers file within 60 days of fiscal year-end, accelerated filers within 75 and everyone else within 90, under the Form 10-K general instructions. The items below carry most of the analysis.

Item	What it holds	Ask AI for
1. Business	Segments, products, customers, competition, human capital	How the company makes money, by segment, in its own words
1A. Risk Factors	Material risks; a summary if the section runs past 15 pages	New, removed and reworded risks versus last year
1C. Cybersecurity	Risk management and governance	Who oversees it, and anything new since last year
3. Legal Proceedings	Material litigation	Each case, its status and any amount accrued in the notes
7. MD&A	Results, liquidity, capital resources, critical estimates	Drivers of each segment’s change, with the figures cited
7A. Market Risk	Rates, FX, commodities	Sensitivities, and how they moved
8. Financial Statements	Statements, footnotes, auditor’s report	Statements across years; footnote tables; critical audit matters
9A. Controls	Management’s and the auditor’s conclusions on internal control	Any material weakness, and the remediation plan
15. Exhibits	Contracts, indentures, subsidiaries, policies	The credit agreement and every amendment

Part III (directors, pay, ownership and related parties) is usually incorporated by reference from the proxy statement, which must be filed within 120 days of year-end. Ask for it from the DEF 14A.

The footnotes that carry the most weight

The statements tell you what happened. The notes tell you how to read them, and they’re where general-purpose AI most often comes up short, because a footnote table rarely matches the words in your question.

Segment note. Revenue and profit by segment, plus a reconciliation to the consolidated totals. Since ASU 2023-07, effective for fiscal years beginning after December 15, 2023, companies disclose the significant expenses their chief operating decision maker sees for each segment (RSM’s summary), so FY2024 was the first 10-K with that detail for calendar-year companies. Ask for segment margins over every year the filings show, and for any segment that was renamed or merged.

Revenue note. Disaggregated revenue and remaining performance obligations, the contracted revenue not yet recognized. For subscription businesses, the change in remaining performance obligations often says more than the revenue line.

Debt note. Each instrument’s principal, rate, maturity and covenants, plus a maturity table. The credit agreement itself sits in Exhibit 10. Ask for every instrument with its terms and page, then have the tool reconcile the total to the balance sheet.

Lease note. Under ASC 842, a maturity table of undiscounted payments for each of the next five years and thereafter, reconciled to the lease liability, plus weighted average terms and discount rates. Kepler reads these tables cell by cell, so a question about year-three lease payments lands on the right row.

Commitments and contingencies. Purchase obligations, guarantees, and litigation. Under ASC 450, losses judged probable and estimable are accrued; reasonably possible losses are disclosed with an estimate or a statement that one can’t be made.

Income taxes. The rate reconciliation, deferred tax assets and the allowances held against them, and uncertain tax positions. ASU 2023-09 expands the rate reconciliation and requires income taxes paid by jurisdiction, starting with public companies’ fiscal years beginning after December 15, 2024.

Note 1. Significant accounting policies, recently adopted standards and the line that matters for anyone building history: “Certain prior period amounts have been reclassified to conform to the current period presentation.”

Read the 10-K against last year’s

The most useful question to ask of a 10-K is what’s different. Five places to look:

1. Risk factors. New risks, deleted risks, and reworded ones. Wording that moves from “could” to “has” usually means the risk has happened, so ask what changed.
2. Accounting policies and critical estimates. A changed revenue recognition policy, useful life or impairment assumption moves earnings without moving the business.
3. Segments. New, renamed or recombined segments, with prior years recast to match.
4. Prior-year figures. Last year’s numbers as printed in this year’s 10-K can differ from last year’s 10-K, and the only notice is often one boilerplate sentence in Note 1. Our guide on as filed vs as restated walks through an Apple example and shows how to track it, and our paper Reclassified to Conform measures how often it happens.
5. The cover page and the auditor. Two check boxes on the 10-K cover flag whether the statements reflect the correction of an error and whether a restatement required a clawback analysis. In the auditor’s report, read the critical audit matters, the judgments the auditor found hardest.

A tool can only answer “what changed” if it has read both filings in full. Kepler holds each company’s 10-Ks side by side and quotes the old and new wording with a link to each.

A 12-question checklist for any 10-K

These work on any tool. In Kepler, the list can be saved as a playbook and run on every 10-K in your coverage as it’s filed.

1. How does the company make money, by segment, and how did each segment’s margin move?
2. Which risk factors are new, deleted or materially reworded since last year?
3. What did management name as the drivers of revenue and margin change in MD&A, and do the statements support it?
4. What is the debt stack: each instrument’s principal, rate, maturity and covenant?
5. What are lease and purchase obligations for each of the next five years?
6. Which contingencies are accrued, which are only disclosed, and what changed?
7. Did any accounting policy, estimate or segment definition change?
8. Were any prior-period figures reclassified or restated, and which lines moved?
9. Is there a material weakness in Item 9A, or an adverse opinion from the auditor on internal control?
10. What do the critical audit matters say about the hardest judgments?
11. Who are the major customers and suppliers, and does any customer exceed 10% of revenue?
12. What’s in the exhibits: new or amended credit agreements, material contracts, related-party arrangements?

Prompts that work

Specific beats general. Ask for a list, the page, and the comparison. Each prompt on the right runs in Kepler as written and returns a cited table.

Weak prompt	Stronger prompt
“Summarize the risks.”	“List every risk factor that is new or materially reworded versus last year’s 10-K. Quote the new and old sentence and give the page of each.”
“What’s the debt?”	“List every debt instrument in the debt note with principal, rate and maturity, cite each line, and reconcile the total to the balance sheet.”
“How are the segments doing?”	“Segment revenue and operating income for every year in the last five 10-Ks, as restated, with any segment renamed or merged flagged.”
“Any red flags?”	“Any material weakness, auditor change, error-correction check box, going-concern language or new related-party transaction in this 10-K or last year’s?”
“Summarize the credit agreement.”	“From Exhibit 10, give the financial covenants, their levels and the definition of Consolidated EBITDA, citing each section.”

Where AI misses on a 10-K

• Footnotes and exhibits. Most assistants retrieve a handful of passages per question. The lease maturity table and a credit agreement that runs well past a hundred pages rarely make the cut.
• Scale. Statement tables are usually in millions or thousands, stated once in a header, and a header that reads “in millions, except per share amounts” means the EPS line in the same table is in dollars. A tool that reads the page has to apply the right scale to nearly every number; our XBRL guide shows how often the printed figure and the stored value differ.
• What changed. Comparing years needs both filings read in full. A model with only one filing in context can describe that filing but can’t see what moved.
• As filed vs as restated. A prior year pulled from this year’s 10-K may not match what was originally reported, and a tool that doesn’t track versions mixes them.
• Conflicts. MD&A text and the statements occasionally disagree. A good tool shows both.

Every answer should pass the checks in how to check AI-generated numbers: open the citation, confirm period, scale and basis, recompute anything derived, and ask twice. For a comparison of tools on filings work, see the best AI for SEC filing analysis.

How Kepler does this

Every 10-K is read in full before you ask. Kepler holds 26M+ SEC filings across 20+ form types, full text and XBRL, for 14,000+ companies. Extraction models we train read every page at ingest, footnote tables, exhibits and scanned pages included, and keep each fact with its value, scale, unit, period and passage. That’s why “every debt instrument” returns every instrument, and why a follow-up question is a lookup rather than another read.

Statements come from the filing’s own XBRL. Code assembles the income statement, balance sheet and cash flow statement from each filing’s tagged facts, with no model in the path. Any line is available as filed or as restated, on any date, and reclassified lines are followed across years with the break marked, using the method in our Reclassified to Conform paper.

The model never writes a number. Frontier models plan the question and write the answer; our deterministic engine pulls every figure and computes every ratio as code over cited inputs. Click a number and land on its line; click a calculated figure and see the formula and each input. When MD&A and the statements disagree, Kepler shows both and flags it.

Your firm’s lens is built in. Our engineers build your firm’s ontology, so your coverage list, your KPI names and your definitions apply to every 10-K. Your 10-K review template is saved as a playbook, and agents run it when a filing lands overnight, with the year-over-year changes waiting in your inbox, in Excel or in Slack. Analysts can run the same review from inside Claude or ChatGPT. Our engineers built this layer before at Palantir, and they set it up with your team on site.

For the data format underneath it all, read what XBRL is and why AI gets statements wrong. Run the checklist above on any 10-K at app.kepler.ai, or book a demo to see it on your coverage.

Frequently asked questions

Can AI analyze a 10-K accurately?

Yes, if it reads the whole filing and doesn't write the numbers itself. Kepler, an enterprise intelligence platform for investment firms, reads every page of every 10-K at ingest, footnotes and exhibits included, builds the statements from the filing's own XBRL with no model in the path, and links every figure to its page and line, as filed or as restated. It also compares each 10-K with the prior year's, so you see what changed.

What are the most important sections of a 10-K?

For most analysts: Item 1 (Business), Item 1A (Risk Factors), Item 7 (MD&A) and Item 8 (Financial Statements and Supplementary Data), especially the footnotes on segments, revenue, debt, leases, income taxes and commitments and contingencies. Item 9A (Controls and Procedures) matters when there's a material weakness, and Item 15 lists the exhibits, where credit agreements and other material contracts are filed.

How do I find what changed in a 10-K since last year?

Compare it with the prior year's 10-K section by section. Look for new, deleted or reworded risk factors, changes to accounting policies and critical estimates, new or renamed segments, prior-year figures that differ from what was originally filed, and the cover-page boxes that flag error corrections. An AI tool can do the comparison if it reads both filings in full and quotes old and new wording side by side.

Can ChatGPT or Claude summarize a 10-K?

Yes. Both summarize long filings well, and their financial services editions connect to filings and fundamentals data. Accuracy on numbers depends on design: a frontier model on its own writes each figure and works from the passages it retrieves, so footnote tables and exhibits can be missed. Check every figure against its line, or use a platform such as Kepler, which runs inside both and returns figures computed from the filing.

What is Item 1C in a 10-K?

Item 1C, Cybersecurity, was added by the SEC's 2023 cybersecurity disclosure rules. It describes how the company assesses and manages cybersecurity risk, the board's oversight and management's role. Material cybersecurity incidents are reported separately, on Form 8-K under Item 1.05, within four business days of the company determining an incident is material.

Which 10-K exhibits should analysts read?

Exhibit 10 (material contracts, often the credit agreement and its amendments), Exhibit 4 (instruments defining security holders' rights, including indentures and the description of securities), Exhibit 21 (subsidiaries), Exhibit 97 (the clawback policy) and Exhibit 19 (insider trading policies). The credit agreement alone can run past a hundred pages and defines the covenants that matter.