TL;DR: A full 10-K averages 150 to 250 pages and most of it is boilerplate. With a structured AI workflow you can extract the five things that actually matter — business model changes, risk-factor deltas, MD&A tone, segment numbers, and accounting footnotes — in under 15 minutes. The trick is knowing what to ask the model, and knowing where it will lie to you.
The 10-K is the most information-dense document a public company produces all year. It is also one of the most boring and most repetitive, which is exactly why people skip the parts that matter. Reading one end-to-end takes three to four hours and almost no one actually does it. What most analysts do instead is open the filing, hit Ctrl+F on a few keywords, read the MD&A header, and close the tab.
That workflow leaves money on the table. The parts of a 10-K that move prices — a single sentence added to the risk factors, a segment footnote that disagrees with the earnings call, a quiet change in revenue recognition — are exactly the parts that keyword search does not find.
Modern AI tools flip the math. Used correctly, they let you extract the signal from a 250-page filing in 15 minutes. Used incorrectly, they give you a confident summary that is subtly wrong in ways that will cost you money. This guide is the workflow we use internally, including the places where we still do not trust the model.
What is actually worth reading in a 10-K
Before touching AI, it helps to know what you are looking for. Of the ten standard 10-K items, five matter for almost every analysis and five are noise most of the time.
Worth your attention:
- Item 1 — Business. Changed segment definitions, new product lines, discontinued operations. A quiet rewrite here usually means the company is repositioning.
- Item 1A — Risk Factors. Only the new risks and the reworded ones. Comparing year-over-year is where the value lives.
- Item 7 — MD&A. Management's own narrative. Tone, hedging language, and the contrast between what they say and what the numbers show.
- Item 8 — Financial Statements. Specifically the footnotes. The income statement tells you what happened; the footnotes tell you how.
- Item 9A — Controls and Procedures. Any disclosed weakness or remediation is a red flag worth understanding in full.
Usually skippable on a first pass:
- Item 2 (Properties), Item 3 (Legal Proceedings, unless new material cases), Item 5 (Market for Registrant's Common Equity), Item 6 (this item is now reserved), and the executive compensation appendices.
If you remember nothing else from this guide, remember this: 95% of the value of reading a 10-K is in the year-over-year delta, not the absolute content. You are looking for what changed.
The 15-minute workflow
Here is the step-by-step sequence. Each step is timed so you can stop when you hit the budget.
Step 1 — Get both filings open (2 minutes)
You need the current 10-K and the previous year's 10-K. Not the earnings release, not the 10-Q — the full annual filings. EDGAR is free and authoritative: sec.gov/edgar/search. Download both as PDFs or grab the direct URLs.
Most AI tools accept either a PDF upload or a URL. URLs are faster because you avoid the download step, but PDFs give you a local copy for later reference.
Step 2 — Run a structured first-pass prompt (3 minutes)
Do not ask "summarize this 10-K." That prompt produces generic output and misses the deltas. Ask something structured instead. This is the prompt template we use:
Compare this 10-K against the previous year's filing. For each of the following sections, list only the changes: (1) business description and segment reporting, (2) risk factors — new risks and removed risks, (3) MD&A tone and forward-looking language, (4) revenue recognition and accounting policies in the footnotes, (5) any disclosed control weaknesses. Do not summarize unchanged content. Use direct quotes for anything material.
This prompt works because it tells the model what not to do (summarize everything) and forces it into a comparison frame. It also demands direct quotes, which is the single most important guardrail against hallucination on financial documents.
Step 3 — Verify the quotes (3 minutes)
This is the step almost everyone skips, and it is the reason people get burned. AI models hallucinate quotes from long documents at a non-trivial rate — current research puts it at roughly 3 to 8% on financial filings depending on the model and document length. A 3% hallucination rate on a 10-K summary means roughly one in every 30 "facts" is fabricated. Some of those fabrications will be material.
Take every direct quote the model returned and Ctrl+F it in the original filing. If the quote is not there verbatim, discard the entire point it supported. This takes about three minutes and it is the difference between a workflow you can trust and a workflow that will eventually blow up a position.
A tool that makes this step easier is worth a lot. Platforms built specifically for filings work usually highlight the source passage when you click a claim in the summary, which cuts verification time roughly in half. NowNews's Reports tool uses this pattern — each answer in the Report Analyst chat links back to the exact passage in the document, so you can click to verify instead of Ctrl+F'ing across 200 pages. We built it this way specifically because we kept getting burned by unverifiable AI summaries and the fix was obvious once we stopped trusting the model by default.
Step 4 — Read the footnotes yourself (4 minutes)
Open the financial statements section and skim the footnotes directly. You are looking for three things:
- Revenue recognition changes. Any new language about how revenue is recognized (timing, bundling, contract assets) is a potential earnings quality issue.
- Non-GAAP reconciliations. Compare the GAAP number to the adjusted number the company highlights in earnings. A widening gap over time is worth understanding.
- Segment reporting changes. If a company re-allocates costs between segments or redefines a segment, the comparisons management shows you in the MD&A are no longer apples-to-apples.
AI is useful here as a "what should I look at" layer. Ask the model: "In the footnotes of this 10-K, flag any accounting policy changes, revenue recognition adjustments, or segment reporting changes relative to the prior year, with direct quotes." Then verify the quotes, per step 3.
Step 5 — Run a sanity check against management's narrative (3 minutes)
The last step is a check against management tone. Read Item 7 (MD&A) and ask the model one more prompt:
Based on the financial statements, are there any claims in the MD&A that are technically true but misleading? For example, management highlighting a metric that has improved while a more relevant metric has deteriorated. List each with the MD&A quote and the contradicting number.
This is where an AI layer with something like sentiment or honesty scoring earns its keep. A tool that just summarizes the MD&A will tell you the CFO is optimistic. A tool that compares language to numbers will tell you the CFO is optimistic about the segment that is shrinking and quiet about the segment that actually matters.
What AI gets wrong on 10-Ks
Three failure modes to know:
Hallucinated quotes. Covered above. Always verify direct quotes against the source. This is non-negotiable.
Averaging away outliers. Language models are trained to produce balanced, fluent summaries. If a 10-K contains one alarming sentence buried in 40 pages of risk factors, a generic summary will often paraphrase it into a softer, averaged version. Structured prompts that ask for "new risks, verbatim" help but do not fully solve this.
Missing the footnotes. Most general-purpose AI tools will extract the income statement numbers but skip the footnotes, which is where the actual accounting decisions live. If a tool cannot show you what it extracted from page 120 onwards, assume it did not read that far.
The practical conclusion is that AI on 10-Ks works best as an acceleration layer, not a replacement layer. It is the difference between reading everything and reading only the things that changed, faster. It is not the difference between reading and not reading.
Tools that handle 10-Ks specifically
General chat interfaces (ChatGPT, Claude, Gemini) handle 10-Ks reasonably well if you feed them the PDF and use structured prompts. Their weakness is verification — you cannot click a claim to see where it came from.
Purpose-built filings tools add that verification layer. In this category the notable options in 2026 are Finchat (conversational research, strong on metrics), AlphaSense (enterprise-focused, expensive, very deep), and NowNews's Reports tool which combines filings analysis with a source-linking chat and is priced at the retail end at €14.99/month for early users. Each has trade-offs; the right pick depends on whether you are reading one filing a week or thirty.
If you want to test a source-linked workflow specifically for 10-Ks, NowNews offers a 7-day free trial without a credit card. Upload a filing, ask the five structured questions from the workflow above, and see whether the verification layer saves you the Ctrl+F time.
Frequently asked questions
How long should it actually take to read a 10-K with AI? With a structured workflow and a purpose-built tool, 15 to 20 minutes for a first pass on a company you already know. Add 10 to 15 minutes if it is a new company and you need to understand the business model. A full deep dive without AI typically takes three to four hours.
Can I just ask ChatGPT to summarize a 10-K? You can, but the output will miss footnote details, may hallucinate quotes, and will not compare year-over-year by default. The summary will feel useful and be partially wrong, which is worse than knowing it is wrong. Structured prompts help; verification against the source is essential.
Which 10-K sections should I never skip? Risk Factors (Item 1A), MD&A (Item 7), and the footnotes in the financial statements (Item 8). Everything else can be skimmed on a first pass. Controls and Procedures (Item 9A) should always be checked for disclosed weaknesses.
How do I spot hallucinations in AI summaries of filings? Every direct quote should be Ctrl+F verifiable in the original document. If a claim does not have a quote, treat it as unverified. If a quote is in the summary but not in the source, discard the entire claim it supported and be suspicious of the rest of the output.
Is it worth paying for a filings-specific AI tool? If you read fewer than five filings a month, general-purpose AI plus manual verification is probably enough. If you read more than that, a tool with source-linked answers will save enough verification time to pay for itself. The main value is not the AI itself — it is the UI that makes verification a click instead of a search.
What about 10-Q filings — does the same workflow apply? Yes, with one adjustment. 10-Qs are shorter (typically 40 to 80 pages) and unaudited, so the workflow compresses to about 8 minutes. The main sections to target are the same, but there are no fresh Risk Factors in most 10-Qs — you mostly compare MD&A tone and footnote changes against the most recent 10-K.
The bottom line
AI does not replace reading a 10-K. It replaces the three hours of skimming with 15 minutes of targeted extraction and 3 minutes of verification. The gain is real, but only if you use structured prompts and check the quotes. If you skip verification, you are making decisions on an output that is confident, fluent, and occasionally fabricated — which is worse than not reading the filing at all.
The workflow above is what we use internally and what the Reports tool in NowNews is built around. If you want to test it on your next filing, start a free trial and upload a 10-K. It takes less time than Ctrl+F'ing through a PDF, which is roughly the point.