The Best Way to Summarize SEC Filings with AI (Without Getting Burned)

A 10-K can be 200 pages. A 10-Q is 40 to 60. An 8-K earnings release plus transcript is another 20 or 30. Multiply that by every company you follow, and keeping up becomes impossible by hand. This is roughly the moment everyone thinks about using AI.

Large language models turn out to be surprisingly good at reading SEC filings. They are not magic, and they will absolutely hallucinate if you let them, but used correctly they can turn an afternoon of reading into a 20-minute review. The trick is knowing where they help, where they do not, and how to structure the workflow so you catch the mistakes before they cost you.

I have been doing this for a while, including building EarningsLens partly because I was tired of the manual process. Here is what actually works.

What AI is genuinely good at

A few specific tasks:

Comparing two versions of the same document. Give an LLM last year's Risk Factors section and this year's, ask it to identify what changed, and the answer is usually accurate. The task is well-bounded, the source text is right in front of the model, and there is no need for the model to know anything beyond what you gave it.

Summarizing a long passage into a shorter one. MD&A sections, earnings call transcripts, footnote disclosures. Give the model the text, ask for a three-sentence summary of a specific section, and it does a creditable job.

Extracting structured data from prose. "Pull out every number mentioned in this paragraph with the unit and what it refers to." This kind of extraction is reliable when the source is clear.

Answering questions where the answer is in the text. "What does management say about gross margin in the Q3 10-Q?" is a question an LLM can answer well if you point it at the Q3 10-Q.

Notice the pattern. All of these are tasks where the source of truth is right there and the model is essentially doing compression, not reasoning about facts it did not see.

What AI is unreliable at

The failure modes are equally predictable:

Anything that requires numbers the model was not given. If you ask an LLM "what was Apple's revenue in Q1 2024" without handing it the filing, you will get an answer that might be right, might be off by a year, and might be entirely fabricated. Model training data is fuzzy on specific numbers. Always give the model the source.

Multi-filing comparisons where you did not hand over all the filings. Asking "how has this company's gross margin trended over the past five years" without giving the model five filings means it is guessing from training data. The answer will sound confident and be often wrong.

Math on big numbers. LLMs are bad at arithmetic, particularly anything involving percentages, growth rates, or multi-step calculations. For any derived metric, calculate it yourself from the numbers the filing provides. The model can tell you the numbers; it cannot reliably divide them.

Value judgments dressed up as facts. "Is this company a good investment" will always get an answer, but the answer is generated from averaged opinion text in the training data. Do not treat it as insight.

Anything relying on very recent news. Filings the model was not trained on, yesterday's 8-K, the announcement that happened this morning. Unless you feed the model the document directly, it does not know.

A workflow that actually works

Here is the pattern I use.

First, decide what you want to know before you ask. Generic "summarize this 10-K" prompts give generic summaries. Specific prompts give specific answers. "Summarize what the MD&A says about the decline in gross margin in Q3, including any quoted explanations from management" is the kind of prompt that works.

Second, always give the model the document. Paste the relevant section in, or use a tool that retrieves the filing and feeds it to the model. Never ask questions about a filing the model cannot see.

Third, ask the model to cite. "Summarize the risk factors related to supply chain and quote the exact sentence for each one" is better than "what are the supply chain risks." The quote lets you verify.

Fourth, spot-check one or two claims every time. Open the filing, search for the quote, confirm it is there and means what the summary says it means. This is the part people skip, and it is the part that separates a useful workflow from a dangerous one.

Fifth, separate extraction from interpretation. Use the AI for extraction (what did management say, what number did they report, what changed). Do your own interpretation (is this good, is it sustainable, does it match what I already believed).

Common mistakes

Treating a confident tone as evidence of correctness. LLMs write fluently regardless of whether the content is right. A clean, professional-sounding summary of a filing can be 90 percent accurate or 60 percent accurate, and the prose quality does not tell you which.

Asking for a single verdict. "Is this a bullish or bearish 10-K" reduces a complex document to a label and encourages the model to pick one. If the filing is genuinely mixed, a forced label hides that.

Using summaries to skip the filing entirely for companies you own meaningfully. For a watchlist of companies you are tracking, AI summaries are a reasonable filter. For a company that is a real position in your portfolio, read the actual filing at least once a year. No summary captures everything, and you do not want to own a company whose 10-K you have never opened.

Letting the model invent numbers. If you ask "what was the company's free cash flow" and the model returns a specific dollar figure, verify it against the cash flow statement before you quote it anywhere. This is probably the single most common hallucination mode.

How EarningsLens handles this

This is what we have been building. A few things that matter:

• Every summary on EarningsLens is generated from the actual filing, which we retrieve directly from SEC EDGAR. The model never answers from memory about a specific filing.
• The Risk Factors diff compares the current 10-K or 10-Q with the most recent prior version from the same company and highlights what changed.
• Key metrics (revenue, EPS, margins) are extracted from the filing as numbers, not interpreted. The numbers you see are what the filing reports.
• The AI commentary is labeled as AI-generated, and each section links back to the original filing so you can spot-check.

It is a tool that speeds you up, not one that replaces your judgment. If you already know what to look for in a filing, it saves you time. If you do not, it can help you learn the shape of what matters, but you should still read a few filings by hand first so you can tell when the summary is missing something.

One honest caveat

Even a well-designed AI tool will occasionally miss things. A subtle footnote change, a buried contingency, an accounting reclassification that matters more than it looks. Professionals who read filings for a living are still more reliable than any current AI. The value of AI is in coverage: it lets you keep a distant eye on thirty companies instead of reading three carefully. For the ones you care about most, you still need to read them yourself.

That combination (AI for breadth, human reading for depth) is the realistic picture of what works today. If someone is selling you on "just use AI, skip the filings," they are either wrong or selling something.