What AI Research Tools Should a Small Hedge Fund Buy First?

Buy in this order: the compliance and recordkeeping baseline, which costs policy time before it costs money; the licensed data the strategy cannot run without; one governed general assistant; the research system that joins licensed data, documents and the fund's own work; and specialist layers last. Licenses come before models because data rights do not transfer. A content license that bars feeding the content to an outside AI tool decides what any model you buy is allowed to read.

We build AllMind, which belongs to the fourth step, so read that section as an interested vendor's case and hold it to the same acceptance test as every other row. Everything here comes from published price pages, regulator texts, vendor terms and one industry survey, opened on September 3, 2026; we tested no product.

What "small" means in this sequence

The 2026 AIMA and Marex emerging-manager survey of 180 managers puts average headcount at 10 and breakeven fund size at $82.9 million, up from $70.1 million in 2024, with 42% of managers deploying AI across all business functions (Marex, June 30, 2026). That is the fund we mean: two to ten people, where the person who signs the contract also uses the tool and answers the examiner.

Headcount never sets the boundary. A two-person concentrated book carries the same records obligation and the same data-rights problem as a fifty-person firm; what changes with size is how much the fund can absorb at once, so the order matters more than the shortlist. Before pricing anything, size the coverage workload in hours so each purchase is tied to the hours it should return.

The sequence, priced

Each row says which kind of figure it carries, with the source we opened; estimates never become list prices.

ChatGPT Business is absent from step 3 because OpenAI's dollar figures render in the browser and our fetcher could not read them. Third Bridge's pricing page shows no figure, and we found none on GLG's site on September 3, 2026, so the expert row is the one per-call price list we could open. Costbench's summary box also prints $2,360 to $2,665 per user per year, one twelfth of the annual figures the row uses. If AlphaSense is the step-four candidate, price the quote with our AlphaSense analysis; model-ready data feeds such as Daloopa sit with the budget lines behind every row in our pricing guide.

Step one costs time before it costs money

Three obligations gate every later purchase, and none is AI-specific. Rule 204-2 requires an adviser's listed records to be "maintained and preserved in an easily accessible place for a period of not less than five years" from the end of the fiscal year of the last entry. Decide with counsel which prompts, outputs and drafts become records, and put the archive in place before the first assistant seat goes live.

The second obligation is what the fund says about its tools. The Marketing Rule bars any advertisement containing "a material statement of fact that the adviser does not have a reasonable basis for believing it will be able to substantiate upon demand by the Commission" (17 CFR 275.206(4)-1). The SEC has charged advisers over false AI claims since 2024; our compliance guide carries the cases, the control set and an acceptable-use outline.

The third is information. Section 204A of the Advisers Act requires written policies and procedures "reasonably designed" to prevent the misuse of material nonpublic information (15 U.S.C. 80b-4a). Expert calls and transcripts, bought in step five, are where a small fund meets that rule first, and a network's own screening does not replace the fund's written procedure. The SEC withdrew its 2023 predictive data analytics proposal on June 12, 2025 (SEC), so there is no AI-specific adviser rule to wait for.

The acceptance test for step one is a mock exam: the AI sentence in the deck, the policy that governs the tool, and the archived record of one research question from last quarter. If all three exist, buy step two.

Licenses before models

The principle behind the order is a clause most funds have never read. AlphaSense's terms of use, updated June 5, 2026, bar using the service or its content with "any third party artificial intelligence or machine learning tool", bar programmatic access, and bar redistributing third-party content beyond the license (AlphaSense terms). Third Bridge's content terms from May 2026 bar forwarding "any Content (or any electronic link to any Content) to any third party" (Third Bridge). So the step-three assistant can read EDGAR and the fund's own notes; the broker note or expert transcript in another vendor's platform stays out of reach unless that license says so, and reading it inside the license is what step four buys.

Broker research has two access models. Aftermarket research arrives on a delay that varies by broker and is included by the platforms that license it; research at publication time runs on the fund's own broker entitlements, paid to the brokers, and a platform connector does not widen that entitlement. Our broker-research licensing guide walks through the contract questions; why the license has to exist before the model reads it is the longer version of this section.

Step two is a question about the strategy. A fundamental long-only or long/short book needs fundamentals, estimates, filings and transcripts under a license the fund holds; Koyfin's Plus plan lists ten years of financials and estimates on its pricing page. A book that trades intraday or lives in Instant Bloomberg messaging needs the terminal seat, and whether that seat belongs in year one is a separate decision.

The free layer is step zero

EDGAR full-text search is free, Koyfin's Free plan is $0, and FRED publishes macro series behind a free account. That layer is enough to write the acceptance tests before any money moves: run the fund's real question, on a real name from its list, through free EDGAR and a free plan, and write down exactly where it fails. Buy for that failure.

What the free layer cannot do is also exact: no license to broker research or expert content, no audit trail an examiner would accept, no permissions model for a second user, and nobody to answer when a number cannot be traced. Each gap is a paid step.

What each purchase unlocks, and when the next one is due

The five-day hedge-fund pilot is the test to run once step four is on the table.

Three fund profiles move the order at the margins

A two-to-four person fundamental fund skips the terminal seat unless someone trades, buys the fundamentals workstation at a published price, adds governed assistant seats for everyone, and decides step four within one earnings season. The trigger is the license wall: if the assistant is asked weekly for something that lives in a broker note or an expert transcript, step four is due early.

A five-to-ten person multi-strategy fund splits step two by book: the trading desk needs the terminal seat, the fundamental book needs fundamentals and estimates, and both share one governed assistant with logging. Step four must enforce per-user permissions because the books are entitled to different content, and that test belongs in the pilot before price is discussed.

A sub-advised or emerging manager on a platform treats step one as a check. The platform may already hold the compliance baseline and some licenses, so the manager confirms in writing which licenses it may use, whose name they are in, and whether platform-held content may enter its own tools. Steps three and four then follow the same tests.

Where we fit in the sequence

AllMind is the strongest first pilot for a small fundamental fund whose first purchase has to replace several subscriptions at once: a filings and transcript search, an estimates feed, aftermarket broker research, an expert transcript library and news.

Licensed data classes sit in one system. Our catalog documents FactSet fundamentals and Revere relationships, LSEG I/B/E/S estimates and revisions, ownership and holdings, live news, and market data across 18 North American and European markets with more than 40 venue feeds (our data catalog). Aftermarket broker research is included on a broker-specific delay, with reports at publication time on the firm's own entitlements. Third Bridge Expert Insights ships as a content class, more than 100,000 expert-interview transcripts with access that varies by package, so a desk reads them without an expert-network contract of its own.

Coverage breadth comes from Grids, which takes a question as a column and researches every cell across the fund's list, and from background agents in Agent Studio that turn a change worth knowing about into an alert. Each answer carries a citation back to the source document, which the step-one archive needs. Permissions and audit close the loop: an agent works inside the entitlements of the person who ran it and nothing wider, access is role-based with SSO, and the log records every access. Our SOC 2 Type II certification dates from November 2025; ISO 27001 and GDPR certification are targeted for Q1 2027.

What the pilot should hand back, before a quote is signed, is a same-day earnings review or a coverage grid over the fund's own list, every material number linked to a filing, transcript, estimate or broker note the fund is licensed to read, archived beside the fund's own notes.

Our limit is how we sell. Pricing is quote-based through a sales form with no self-serve checkout, so a fund that wants to start this afternoon with a credit card cannot start with us; a published-price tool starts today, and the sequence still holds. Two cases send the first purchase elsewhere. Live market data and execution are a terminal purchase and belong in step 2b. A fund that only needs a governed chat assistant over EDGAR and its own notes should buy that seat, run step three's test for a quarter, and revisit step four when the license wall appears.

What stayed unverified

Every quote-only price is unverified, including our own, AlphaSense's list price (Vendr's figure is a deal sample), S&P Capital IQ Pro, Hebbia, Rogo, Third Bridge, GLG, Quartr, Aiera and Visible Alpha. Bloomberg's own Terminal pages returned 403 to every fetch we tried on September 3, 2026, so row 2b rests on two third-party estimates. The Fiscal.ai and Perplexity Enterprise pricing pages blocked every fetch method we tried. We found no SEC AI-washing action against an adviser dated 2026; the cases we know of run through 2025. No survey we opened publishes a technology-spend figure for funds under $100 million.

Frequently asked questions

What is the best AI research stack for a fundamental hedge fund?

It does not change with headcount: the layers in the sequence table are the same for two people or ten. What moves is the purchase order and the date step four falls due, which the three fund profiles above set by one trigger, the first time the assistant is asked for content inside another vendor's license.

Is a Bloomberg Terminal overkill for a small hedge fund?

It depends on the job the seat does. If someone trades, needs live pricing or works inside Instant Bloomberg messaging, the seat does a job no research tool replaces and belongs in step two. If the seat only reads filings, pulls estimates and searches research, a published-price workstation plus a research system covers that work for a fraction of the terminal estimate in the table above.

Can a small fund put licensed broker research or expert transcripts into ChatGPT or Claude?

Only if the content license allows it, and the licenses we opened do not. AlphaSense's terms bar using its content with any third-party AI or machine-learning tool, and Third Bridge's content terms bar forwarding content or links to any third party. Public filings and the fund's own notes can go into a governed assistant; licensed content is read inside the platform that holds the license, or under written permission from the licensor.

Sources and methodology

Nothing here is legal advice.

• Regulatory texts: Cornell LII copies of Rule 204-2, Rule 206(4)-1 and Advisers Act Section 204A; the SEC's June 12, 2025 withdrawal notice for S7-12-23. Opened September 3, 2026.
• Published prices: Koyfin, Claude and the AlphaSense expert-call page (its average call cost is vendor-reported), as read on September 3, 2026; prices change without notice.
• Third-party estimates: Costbench, Bloomberg Terminal (updated August 22, 2026, 15 purchases); Wall Street Prep, older terminal figures (updated April 7, 2025); Vendr, AlphaSense (February 2026, 38 purchases).
• License terms: AlphaSense terms of use (June 5, 2026); Third Bridge content terms (May 2026).
• Small-fund profile: AIMA and Marex Emerging Manager Survey 2026 (June 30, 2026).
• Absence checks: Third Bridge's pricing page, GLG's site and Bloomberg's Terminal pages, tried September 3, 2026.
• AllMind statements are our own, from our data catalog, product pages and pricing page as read on September 3, 2026.

If the fund is at step four, bring the coverage list, one restricted document and last quarter's earnings review to the pilot. Ask us and every other candidate for the same three things in the same week: the finished artifact with its citations, the failed permission test, and an itemized quote.