Can an AI assistant produce SEC filing research that holds up in an investment-committee memo? Yes — if every number it writes is cited to a source filing, and the same question returns the same evidence twice. In the session below, Claude — connected to SEC data through the edgar.tools MCP server — split e.l.f. Beauty’s 25% FY26 revenue growth into $293.5 million from the rhode acquisition and $29.5 million from the existing business, quoting the exact MD&A paragraph of a 10-K filed three days earlier and citing it to SEC accession 0001600033-26-000020. Nobody typed the word “rhode.” Time from question to cited answer: under two minutes. This post shows how the session got there, why the output is auditable enough for a memo, the time math against doing it by hand, and what it costs next to the tools a fund already pays for.
A note on who’s talking: I build edgar.tools. I’m not an analyst — so the workflow below is the one analysts describe to me, run end to end on live filings, with the receipts kept.
How do hedge fund analysts use AI for SEC filings?
The short answer: ask plain-English questions in Claude, get answers grounded in structured filing data, with a click-through SEC citation on every number. No new terminal, no query language — the tools load into the Claude Desktop your team already uses.
Here’s the scenario, stated honestly as a scenario: it’s Sunday night, the IC meets at 9:00 a.m., and one of the names to prep is e.l.f. Beauty — on the watchlist because revenue has been climbing fast and the stock has had a wild ride. The persona for the session is a value-side investor with thirty minutes. Six questions, eight tool calls, and this is what came back:
- Headline financials, from a filing three days old. FY26 (ended March 31, 2026): revenue $1,636.5M, net income $26.3M, operating cash flow $212.5M — each figure carrying the accession of the 10-K it came from.
- The trend inflection, pre-computed. Net-income growth swung from +107.5% in FY24 to −12.2% in FY25; operating-income growth decelerated from +120% to +5.6%. Claude didn’t do that arithmetic — the tool returned the inflection as a flagged signal, each datapoint cited to its source filing.
- The ownership tape. Sixteen insider transactions in 180 days, and the one that answered the real question — anyone buying without being made to? — arrived as a single structured field: a director’s open-market purchase of 5,000 shares at $92.96 on 2026-02-20, purchase code P, no compensatory footnotes. Institutional side: 456 holders, $4.78B reported value, Baillie Gifford on top at $545.8M, from the December 31, 2025 13F cycle.
- The new risks. A risk-factor diff between the FY25 and FY26 10-Ks: two added, seventeen materially changed. The top added entry: “Acquisitions or investments, such as our acquisition of HRBeauty LLC (‘rhode’), could disrupt our business and harm our financial condition.”
- The kicker. Asked “what drove FY26 revenue growth?”, Claude carried the 10-K accession forward from the risk diff and pulled the MD&A paragraph: net sales up $323.0M (+25%) to $1,636.5M — $293.5M contributed by the rhode acquisition, $29.5M from the existing business.
Run the arithmetic on that last line and the existing business grew a little over 2%. That is the kind of decomposition an analyst would otherwise have produced with an afternoon and a highlighter — and to be clear, it’s not a view on the stock. It’s what the filings say, with the page they say it on.
Can you trust AI output in an investment memo?
You can trust what you can audit, and that is the design constraint everything here was built around. A memo claim that traces to an SEC accession number is checkable by anyone on the committee; “the AI said so” is not.
Four properties do the work:
- Every claim is cited. Each number above carries the accession of the filing it came from — one click to the source document on sec.gov.
- Structured data underneath, not scraped prose. Every response names its source filing and as-of date, and the quantitative work (trend math, diff counts, transaction-code classification) happens in deterministic tools, not in the model’s head. The FY25 deceleration arrived pre-computed and flagged; Claude’s job was to write it up.
- It says nothing when there’s nothing. Run the same thesis workflow on NVIDIA over a 90-day window and the risk section reads: no material change — 0 added, 0 removed, 0 materially changed between accessions 0001045810-25-000023 and 0001045810-26-000021, with the confidence line explicitly flagging the absence of insider Form 4 activity as a limit on the read. A tool that can say “nothing happened, here are the two documents I compared” is the same tool you can believe when it says seventeen risk factors changed.
- It refuses what isn’t in filings. No price targets, no analyst consensus, no buy/sell calls — those don’t live in SEC documents, and the toolkit won’t invent them. The refusal is a feature: it marks the boundary of what’s grounded.
And the evidence layer is repeatable: the same question against the same window assembles the same filings and cites the same accessions. Only the prose varies — the audit trail doesn’t.
How long does a thesis brief take with AI vs. manually?
The manual baseline, as analysts describe it: a fresh workup on a long or short candidate means pulling the 8-K item codes, the risk-factor diff against the prior 10-K, two quarters of 13F flow, the Form 4 tape, and a peer comparison. Five lookups across four tools — call it an hour of clicking.
The same workup as a slash command — /edgar:position_thesis ticker=NVDA lookback=90 — assembles all of it and returns the structured brief in about 65 seconds. The full ELF question-and-answer session above, including the time spent actually reading the answers, fit inside thirty minutes.
It compounds across a book. A four-ticker BDC sweep (ARCC, MAIN, BXSL, PSEC) came back from a single call in about 12 seconds, clustered by theme rather than by date — and ranked Prospect Capital’s combined Items 1.01/3.03/5.03 filing (a material agreement plus a modification of security-holder rights plus a charter amendment, in one 8-K) at the top by materiality, above later-dated earnings releases. During earnings season, that triage is the difference between reading everything and reading what matters.
The pre-built workflows on the Analyst tier:
| Slash command | What it returns |
|---|---|
position_thesis |
Setup, catalysts, risk deltas, ownership signal, view + confidence — one name |
portfolio_pulse |
Themes, top-3 material events, quiet names — up to 50 tickers in one call |
peer_comparison |
Aligned peer table + spread narrative |
earnings_postmortem |
Beats/misses, guidance changes, red flags from the latest print |
filing_red_flags |
Ranked adverse signals with cluster detection |
What does it cost compared to AlphaSense or a Bloomberg terminal?
A terminal seat runs on the order of $30,000 a year; enterprise research platforms price annually, per seat, in the five figures. edgar.tools is $24.99/month for the Pro toolkit (profiles, trends, ownership, filing narrative, corpus search) and $79.99/month for Analyst, which adds the intel bundle and the slash-command workflows above.
This isn’t a terminal replacement — there’s no market data, no chat, no execution. It’s the SEC-filings layer of the research stack, priced like software instead of like an institution. The rest of the math: self-serve signup, a 14-day trial, no seat minimums, no sales call, and nothing for IT to deploy — it runs inside the Claude your team already has. If it saves one analyst-afternoon a month, it has covered its cost several times over.
Bring a name from your book
Setup is one install: edgar.tools runs an MCP server (the open standard for connecting Claude to external tools — that’s the last you’ll hear of the acronym). Install from the Plugin Hub, authorize with your account, and the tools appear in the same session. The five-minute walkthrough is in the setup guide.
Then run the test that convinced me this post was worth writing. Pick a name from your own coverage — one where you already know the story — and ask:
“What drove [company]'s revenue growth last fiscal year, and where in the 10-K is that documented?”
If what comes back matches what you know, with the citation already on the line, you’ve just watched your Sunday night get shorter.
Start a 14-day Analyst trial →
Dwight Gunning is the creator of edgartools, the open-source Python library for SEC data, and builds edgar.tools. The ELF and NVIDIA sessions above were run against live filings in May 2026; every accession cited resolves on sec.gov.