TL;DR
A 10-K runs 100-300 pages and most of it is boilerplate. To get a usable one-page brief: pull the official filing from SEC EDGAR, feed the PDF to a long-context model (Claude or NotebookLM both hold a 150-page filing without chunking), run the prompt below, then verify the three headline numbers against the source before you forward anything. AI is excellent at compression and pattern-spotting; it is unreliable at reading numbers off tables, so the verify step is not optional.
The task
You need to read an annual report fast — for competitive analysis, due diligence, a partnership review, or because a manager dropped it on your desk. The full PDF is 100-300 pages, and roughly 60% of it is risk-factor boilerplate, legal disclaimers, and standardized financial-statement tables. You want a one-page brief: the business model, revenue breakdown by segment, the three biggest risks, the single biggest opportunity, and the gap between the line the CEO is proudest of and the line they buried. The brief is for you, not the regulator — write for clarity, not compliance.
Get the authoritative file first
Don’t summarise a PDF you found on a random investor-relations blog. Pull the original from SEC EDGAR, the SEC’s free public database — no registration, full-text search, and machine-readable XBRL financial statements for every US public company back to 1994. Search by company name or ticker, filter form type to 10-K, and download the latest one. For the year-over-year comparison you’ll want the prior filing too. XBRL data sets let you cross-check the headline figures the AI reports.
When AI helps — and when it does not
AI is strong at compressing dense accounting language into plain English and surfacing structural patterns: segment growth, margin compression, a geographic shift in revenue, a new risk factor that wasn’t in last year’s filing. It is weak at reading numbers off tables. Models still misalign rows and columns and then state a wrong revenue figure with full confidence. MIT researchers found that models use more confident language when they are wrong than when they are right, which is exactly why a misread number is easy to miss. A documented failure mode: ask for “revenue growth from $50M to $30M” and a model may answer “50%” instead of −40%. Treat every figure as a draft until you check it.
| Task | AI reliability | What to do |
|---|---|---|
| Plain-English business model | High | Trust, skim once |
| Segment / risk-factor narrative | High | Spot-check against headings |
| Headline revenue, growth, margin | Low — verify | Read the income statement yourself |
| Growth-rate / ratio math | Low — verify | Recompute by hand |
| Footnote-derived claims | Low — verify | Open the footnote |
Which tool for a 150-page filing
| Tool (June 2026) | Long-doc handling | Best for | Price |
|---|---|---|---|
| Claude Pro (Sonnet 4.6) | 1M-token context, holds whole filing; PDFs over 100 pages get text-only extraction | Synthesis and reasoning across one report | $20/mo |
| NotebookLM | Strict inline numbered citations to your source | Grounded extraction; querying several reports at once | Free / in Google AI Pro $19.99/mo |
| ChatGPT Plus (GPT-5.5) | 512 MB/file upload; ~320-page in-app context; slows and formats poorly past ~900 pages | Quick single-section questions | $20/mo |
| Gemini 3.1 Pro | 1M-token context; 100 MB/file | Cross-source synthesis | Google AI Pro $19.99/mo |
A practical division of labour from equity-research practitioners: NotebookLM for extraction (it forces a citation on every claim, so you can click straight to the source line), Claude for synthesis (it reasons better across the whole document). If you only use one, Claude Pro at $20/mo handles a single 150-page 10-K end to end.
What to feed the AI
- The 10-K / annual report (PDF from EDGAR, or the specific sections pasted in)
- The reader’s lens (competitive analyst, non-finance exec, partner, investor)
- Specific questions you already have (“what is happening in the international segment?”)
- Industry context the model cannot infer (a recent regulatory change, a peer benchmark)
- What you’ll use the summary for (briefing, deck, investment memo)
Copy-ready prompt
Summarise this 10-K / annual report into a one-page brief.
Reader lens: [competitive analyst / non-finance exec / partner / investor]
Specific questions I have: [list]
Industry context you should know: [list]
Use case: [briefing / deck / memo]
Document:
"""
[paste relevant sections or whole report]
"""
Return:
1. Business model in 2 sentences (no jargon)
2. Revenue breakdown by segment, with YoY change
3. Three biggest risks (from the MD&A or Risk Factors section)
4. The single biggest opportunity, in plain language
5. One thing the CEO is most proud of (with the quote)
6. One thing they soft-pedalled (with the line that hints at it)
7. The most underemphasized risk — risks get buried near the end of the section
8. Three follow-up questions an analyst would ask after reading
Do not paraphrase legal disclaimers. Mark every number you report with
[VERIFY], and tell me which page each one came from so I can check it.For the comparison pass, add the prior-year filing and ask: “Compare against last year’s report — what’s new, what’s gone, what changed quietly, and which risk factor was added or removed.”
How to check the output is usable
- Revenue, growth rate, and operating margin match the income statement you read yourself
- The business model fits in 2 sentences a stranger can understand
- Risks are real items from the Risk Factors section, not generic placeholders
- The “soft-pedalled” item is a genuine hint from the report, not editorialising
- Every reported number has a page reference you can click to
- The brief earns your reader’s attention without making them re-read the source
Common mistakes
- Skipping the risk section. Annual reports bury bad news in Item 1A and in the back of the MD&A — that’s where the real story is.
- Trusting the model’s number recall. Verify headline metrics against the income statement; never quote a figure from memory.
- Letting AI invent customer-concentration or churn figures. Only use them if they’re actually disclosed in the filing.
- One-shot summary with no prior-year comparison. The year-over-year change is the signal; a single year is just a snapshot.
- Forwarding to leadership before verifying. One wrong number ruins your credibility on the whole brief.
FAQ
Long PDF — chunk it or feed the whole thing? With a 1M-token model (Claude Sonnet 4.6, Gemini 3.1 Pro) a single 150-page filing fits whole. For 200-300 pages, or for cross-report queries, chunk by section or use NotebookLM, which is built to query large source sets and cites each claim.
Can AI read tables and footnotes accurately? It reads them, but accuracy drops on dense tables — row/column misalignment is the classic error. Note that Claude only parses images and charts inside PDFs under 100 pages; over 100 pages it extracts text only. Verify any footnote- or table-derived claim against the page.
What if the company is private and has no 10-K? Swap in the materials you do have — investor updates, board decks, audited statements, a 409A valuation report. The same prompt structure works; just tell the model what kind of document it’s reading.
Which model is most accurate on the numbers? None are reliable enough to quote without checking. As of June 2026, top models still misstate figures from tables, and they do it confidently. Use NotebookLM’s citations or Claude’s page references to jump to the source, and read the headline numbers off the statement yourself.
Related
- PDF summary prompts — broader PDF summary workflow
- Competitor comparison — when the summary feeds competitive analysis
- Competitor analysis AI — structured competitor teardown
- Business data analysis AI — analyse business data adjacent to the report
- Market research organize AI — broader market research
- Chart table explanation AI — make tables digestible
- AI fact check workflow — verify quoted claims