How Hedge Funds Use AI for Research in 2026

Hedge funds use AI for research mostly to read faster and draft sooner: screening for ideas, summarizing filings and earnings calls, updating models when results land, tracking theses and monitoring news on the book. Kepler, which we build, handles the three places where generic AI still costs analysts hours of checking (numbers, complete screens and the fund’s own definitions): our engineers build each fund’s own ontology, extraction models read every filing in full before a question is asked, and code computes and cites every number.

This guide follows a fundamental analyst through a typical week and shows where AI helps, where it breaks, and what the largest funds have learned. It’s written for long/short equity, credit and multi-strategy research teams; quant signal research is a different topic. We build Kepler, enterprise intelligence for investment firms, so each step below also shows how Kepler handles it, and the last section shows how our agents, our calculation engine and each fund’s own ontology fit together. Facts are as of November 2026.

How widely hedge funds use AI

Adoption is broad, and the most advanced funds have moved from chat to systems.

• Industry surveys. The Alternative Investment Management Association’s 2025 report, Charting the course: Lessons from AI leaders in alternative investment, surveyed 150 fund managers with about $788 billion in assets under management, plus 18 large institutional investors. It’s AIMA’s second study of generative AI at hedge funds, and the AI leaders it interviewed argue that the AI era “will not reward those who are overly cautious.”
• A multi-strategy fund’s build. In a March 2026 case study, OpenAI describes the AI research system at Balyasny Asset Management, built by an applied AI team of about 20 and used by about 95% of its roughly 180 investment teams.
• A macro fund’s assistant. At Claude for Financial Services’ launch in July 2025, Bridgewater’s AIA Labs said Claude powered the first versions of its Investment Analyst Assistant, which writes Python, builds charts and works through financial analysis tasks.
• Vendors moving in. OpenAI launched ChatGPT for Financial Services on September 10, 2026, Bloomberg made its ASKB assistant generally available on October 15, 2026, and Anthropic has expanded Claude for Financial Services with data connectors and Office add-ins.

The pattern: the largest funds build their own systems on frontier models, and most funds pair a frontier assistant with a specialist research platform. Kepler is built for that second group, and because it runs inside ChatGPT and Claude, a fund’s analysts can stay in the assistant they already use while the numbers come back computed in code. For a vendor-by-vendor view, see the best AI tools for hedge funds.

What the leading funds do differently

The published accounts from funds furthest along share four habits, set out most clearly in OpenAI’s Balyasny case study:

1. They test models before deploying them. Balyasny scores models on more than 12 dimensions against its own benchmarks and data, then picks models task by task.
2. They build feedback loops into the work. Feedback from merger arbitrage teams showed that agents needed to re-evaluate deal probabilities continuously as new filings and press releases came in, and the AI team built that into the agents.
3. They centralize the platform and customize locally. A central team builds the agent frameworks, tools and compliance guardrails, and each investment team gets scoped access to the data and tools its strategy needs.
4. They keep people on the decision. Analysts use AI to get to structured, explainable inputs faster; the call stays with the PM.

Most funds can’t staff a 20-person team to build all of that, and Kepler gives them the same four habits ready-made. We test models task by task, as Balyasny does, route each task to the one that performs best, and replace it when a better one ships. Agents re-run when a new filing or press release lands, and the corrections analysts make become labels that improve the fund’s own models privately. One platform serves the whole fund, with permissions, definitions and playbooks scoped to each team. And people decide: nothing leaves the fund or gets shared without a person.

The analyst’s week, and where AI fits

Workflow	Where AI saves time	Where it breaks	How Kepler handles it
Idea generation and screens	Reading thousands of filings for a theme	Screens need complete, consistent numbers	Every filing read in full; screens computed by code across the universe
Earnings season	Transcript summaries, guidance comparisons, first-draft notes	Non-GAAP vs GAAP, restated prior periods, numbers in the note	Model updated within the hour, every changed cell cited
Model maintenance	Spreading a new quarter into the model	Wrong period or units, inputs typed by the model	Statements from XBRL, as filed or as restated, scale and units checked
Thesis tracking	Alerts on filings, news and guidance	Alerts without the exact quote, disclosures missed	Agents on the book; alerts quote the exact sentence and filing
Alternative data	Cleaning, tagging, matching to tickers	Entity resolution, point-in-time history	Licensed datasets resolved into your ontology, point-in-time
Compliance and records	Logging, review support	MNPI, recordkeeping, claims about AI use	One permission check on every read; every query and agent run logged

Idea generation and screens

The useful part is reach. An analyst can ask which companies mentioned a new tariff exposure, a supplier change or a pricing action across a whole sector’s filings and calls, and get a reading list in minutes. The weak part is the quantitative screen. “Retailers whose inventory grew faster than sales for three straight quarters” needs every company’s inventory and revenue, for the right quarters, on a consistent basis. A general model searches, picks the strongest matches and computes what it found, which gives a list that looks complete and isn’t guaranteed to be. Kepler runs the same screen as code over figures our extraction models pulled from every 10-Q at ingest, on one basis, so the list covers the whole sector and each quarter opens on its filing line. Our guide to screening companies with AI shows how to check that a screen is complete.

Earnings season

This is where most funds feel AI first. Transcripts get summarized within minutes of a call, guidance gets compared with the prior quarter and consensus, and first-draft notes are ready before the analyst has finished listening. What still needs care is the arithmetic in the note: adjusted against reported figures, prior periods the company has recast, and growth rates computed off the wrong base. During a heavy week, with a dozen names reporting across two days, those small errors compound, because nobody has time to trace every figure before the morning meeting. Kepler’s agents update the fund’s model within the hour a 10-Q or earnings release lands, cite every changed cell, keep adjusted and reported figures apart, and draft the post-print note in the PM’s format. Our guide to AI for earnings season walks the week step by step, and our list of the best AI for earnings call analysis compares tools on exactly this.

Model maintenance

Rolling a model forward a quarter is tedious and error-prone, which makes it a natural target. Excel add-ins from Anthropic, OpenAI and Microsoft now build and edit workbooks directly. The risk sits in the inputs: when a general model types historical figures into the model, each one inherits the model’s error rate. On Vals AI’s Finance Agent benchmark, updated October 29, 2026, the best overall score was 67.85% and the best Financial Modeling score 57.0%. In Kepler, no model types an input: statements are built from the XBRL in each filing, as filed or as restated, with scale and units checked, and the new quarter lands in the fund’s own workbook in Excel.

Thesis tracking and monitoring

Analysts want to know the moment something changes for a name they own: guidance, a covenant, a related-party disclosure, a customer concentration. AI can watch more sources than any person. The test is whether an alert carries the exact sentence and filing, so the PM can act on it, and whether the system read the whole document or only the passages a search returned. Kepler’s agents read every new filing for every name on the book and the watchlist, and each alert quotes the exact sentence with a link to the filing, in Slack or email.

Alternative data

AI helps clean and tag messy datasets such as web traffic, job postings and app downloads, and match them to companies. Entity resolution is the hard part: subsidiaries, brands and renamed companies have to map to the right ticker, and for backtests the history has to be point-in-time, as it was known on each date. Kepler resolves subsidiaries, brands and renamed companies to one record in the fund’s ontology, and every fact keeps the date it became public, so a backtest sees only what was known at the time.

Compliance and recordkeeping

AI raises old questions in new places. Prompts and outputs can contain material non-public information, research records have to be kept, and what a fund tells investors about its AI has to be true: in March 2024 the SEC settled charges against two advisers for false and misleading statements about their use of AI, with $400,000 in combined penalties. Our guide to AI governance for investment firms covers what to document before an exam. In Kepler, one permission check that mirrors the fund’s walls applies to every read, by a person or an agent, and every query, calculation and agent run is logged with its lineage, so compliance can show where each figure came from.

Where AI breaks in hedge fund research

Across the week, the failures cluster in a few places:

• Numbers written as text. A general model generates figures, so one wrong number reads like the right ones.
• Top hits instead of everything. Screens and disclosure searches return what retrieval found.
• Someone else’s definition. The model applies a generic adjusted EBITDA or net debt unless told otherwise, every time.
• Mixed versions. As-filed and restated figures blend, which quietly breaks a backtest.
• Documents that give instructions. A filing, email or web page can contain text that tries to steer the model.
• Walls and permissions. An assistant that sees everything can put restricted information in front of the wrong person.

Kepler is designed around each of these. Code computes every number, every document is read in full before anyone asks, the fund’s definitions live in its ontology, every figure is labeled as filed or as restated, nothing Kepler reads is taken as an instruction, and one permission check mirrors the fund’s walls.

Build or buy

Balyasny’s 20-person applied AI team is the build model at full scale: evaluation pipelines, agent frameworks, compliance guardrails and a central platform that each investment team adapts. Few funds can staff that. Most buy a research platform and point their own engineers at the parts that make the fund different, such as signals, house views and workflows. The question to ask a vendor is which layers it gives you: the data, the entity model, the calculation engine, the citations, the agents and the controls. Our guide to build vs buy works through that decision, and our best AI tools for equity research list covers the main options.

That’s the gap we built Kepler for, and it supplies all six layers. Building our data layer took 17 months, 109 releases and more than 700,000 lines of code. Funds with engineering teams buy that layer and build their own signals and workflows on Kepler’s data, ontology and engine through our API and Python SDK, instead of rebuilding it first.

How Kepler does this

We build Kepler.

We designed Kepler for the analyst who owns a book and has to be right about it. In the first week, our forward-deployed engineers build your fund’s ontology: the book, the watchlist, each name’s peers, your house KPIs and your definitions, with every company, subsidiary and brand resolved to one record and every link tied to its source sentence. Public data is already loaded, 26M+ SEC filings and 40M+ IR documents and earnings call transcripts for 14,000+ companies in 120+ countries, and extraction models we train have read all of it in full. Your research notes, models, email and Slack connect read-only and get read the same way.

Code computes every number, as filed or as restated on any date, and the same question returns the same figure. Click any number to open the page, line or cell behind it; calculated figures open with their formula and inputs. Each citation is matched against its source before it reaches you, and where two sources disagree you see both, with the conflict flagged. Each step goes to the model best suited to it: small models we train handle the reading, frontier models the question and the write-up, so each new frontier model makes the fund’s setup better.

Agents carry the routine work: overnight runs on the book, model updates when filings land, guidance alerts and a digest before the morning meeting, each run recorded and each acting as the analyst who set it up. Your PM’s earnings-update format and your screen definitions are saved once as playbooks, and the corrections your analysts make become labels that improve the fund’s own models privately.

Every read passes one permission check that mirrors your walls, every query and agent run is logged with its lineage, and text inside a filing or email can’t redirect an agent, because nothing Kepler reads counts as an instruction. We’re SOC 2 Type II, and the fund chooses our cloud or its own. Analysts use Kepler in Excel, Slack, Claude and ChatGPT, and your engineers build the fund’s own signals on it through our API. Our engineers built this layer before at Palantir, and CEO Vinoo Ganesh was Citadel’s first Head of Business Engineering.

• “Alert me the moment any holding cuts full-year guidance, with the sentence and the filing.”
• “Software names in our coverage where net revenue retention fell two quarters running, each quarter cited to its filing or call.”
• “Retailers whose inventory grew faster than sales for three straight quarters, with every input cited.”

See Kepler for finance, try it at app.kepler.ai, or book a demo and we’ll run it on your book.

Frequently asked questions

How do hedge funds use AI?

Fundamental hedge funds use AI mostly to read faster and draft sooner: screening for ideas, summarizing filings and earnings calls, updating models when results land, tracking theses and news on the book, and cleaning alternative data. Kepler, which we build, does that work on each fund's own ontology: every filing read in full before anyone asks, every number computed and cited in code, and agents watching the book overnight, so screens come back complete and figures go straight into the model.

Do hedge funds use ChatGPT or Claude?

Many do, usually through enterprise plans or the API rather than consumer apps. OpenAI published a case study on a multi-strategy fund whose AI research system is built partly on OpenAI models, and Anthropic has said Claude powered the first versions of a research assistant at Bridgewater's AIA Labs. Kepler runs inside both assistants.

Will AI replace hedge fund analysts?

Not on current evidence. On Vals AI's Finance Agent benchmark, which tests entry-level analyst work on public filings, the best model scored 67.85% in October 2026, and the top Financial Modeling score was 57.0%. Funds are using AI to take reading, first drafts and monitoring off analysts' desks. With Kepler, agents do that work overnight and every figure arrives computed in code and cited, so analysts spend more of the week on judgment and the variant view.

What are the risks of using AI in hedge fund research?

Wrong numbers that look right, incomplete screens, the model applying its own definition of a metric instead of the fund's, mixing as-filed and restated figures in a backtest, documents that try to steer the model, and gaps in records for compliance. The SEC has also fined advisers for overstating their use of AI, so claims to investors need to match reality. Kepler is designed around each of these risks, from numbers computed in code to one permission check that mirrors the fund's walls.

Should a hedge fund build or buy AI research tools?

The largest funds build: one multi-strategy firm's applied AI team has about 20 people. Most funds can't staff that, so they buy a platform and build on top of it. Building our own data layer took 17 months; with Kepler, a fund's engineers get the data, the ontology and the calculation engine ready-made, and spend their time on the fund's edge instead of rebuilding infrastructure.