Blog · · Ashutosh Agarwal
Research: How AI Has Changed the Hedge Fund Analyst Due Diligence Workflow
AI has not removed hedge fund due diligence. It moved the bottleneck from finding documents to proving which evidence matters. Here is what changed.
TL;DR
AI has not removed hedge fund due diligence. It moved the bottleneck. Finding and reading the corpus used to be the constraint. Now an analyst can interrogate a decade of filings, transcripts, peer commentary, supplier read-throughs, expert notes, podcasts and web data in one pass, and the hard part is deciding which evidence survives into the memo.
The traditional equity research due diligence process asked an analyst to process the corpus by hand: read the latest filings, skim the relevant transcripts, pull the sell-side model, book two or three expert calls, and fold the assumptions into their own model. That process still describes what most funds do. What has changed is that the reading is no longer the job. Orchestrating the agents that read, and judging what they surface, is the job.
This is the practical shape of AI hedge fund research today, and it is worth being precise about what actually moved in the buy-side research workflow, because most of the noise about generative AI at hedge funds skips that part.
heading: In this article
What is the hedge fund due diligence workflow?
Why one name now means a thousand documents
What changed across the industry in 2025 and 2026
The old diligence workflow next to the new one
Where the bottleneck actually moved
Why agentic collection beats a traditional scraper
Running diligence as a chain of playbooks
Diligence on a name is diligence on its ecosystem
What AI does not change
What is the hedge fund due diligence workflow?
The hedge fund due diligence workflow is the sequence a fundamental analyst runs between finding a name and pitching it: read the filing history, build the model, work the earnings call transcripts, read sell-side research, test the thesis against alternative data, call experts to close the remaining gaps, then write the memo. At a discretionary long/short fund it typically takes four to six weeks per name and produces one output that matters, which is a defensible view of what the market has wrong.
Fundamental long/short analysts pride themselves on the depth of that process. Managers construct their own view of the world and build a nuanced mental model for every company they cover. They leave no stone unturned. They call that information advantage "edge", and one of the first questions a PM asks is what our edge is on this name. The answer had better be compelling.
Why one name now means a thousand documents
The sources that produce edge have multiplied, and the frequency has gone up with them. To understand one public company properly, an analyst reviews the 10-Ks and 10-Qs and captures hundreds of KPIs, all of which have to be maintained in a model every quarter or more often. Then the call transcripts, prepared remarks against Q&A, how guidance language shifted over time, how the market reacted, and the same exercise again for competitors, suppliers and customers. Then sell-side research, alternative data, web traffic, scraped pricing, app data.
One name over ten years is ten annual reports, thirty quarterly reports, forty transcripts, a decade of proxies, investor days, 8-Ks, conference presentations and model history. Add five competitors and three suppliers and the file runs to five hundred or a thousand primary documents before a single broker note.
The documents also keep getting heavier. Dyer, Lang and Stice-Lawrence, in the Journal of Accounting and Economics, documented that across roughly 76,000 firm-years from 1996 to 2013, 10-K textual disclosure got longer, more boilerplate, more redundant and less readable, with three topics (fair value, internal controls and risk factors) accounting for almost all of the growth. The SEC's EDGAR full-text search will hand you every one of those filings in seconds. Reading them was never the part a search box solved.
What changed across the industry in 2025 and 2026
Adoption stopped being the question. AIMA's September 2025 report surveyed 150 fund managers representing about $788 billion in AUM and put generative AI use at 95%, up from 86% two years earlier, with 58% expecting to increase front-office use inside a year.
The results are showing up in the workflow itself. In OpenAI's case study on Balyasny Asset Management, about 95% of investment teams work through an internal AI research platform, deep research tasks that used to take days now take hours, and one specialist agent cut a macro scenario analysis from two days to roughly thirty minutes. Balyasny staffed a central applied AI group of about twenty researchers and engineers to build those tools into team workflows.
Alternative data went the same way. Lowenstein Sandler's 2025 alt data survey found adoption up from 62% of funds in 2023 to roughly 90%, which turned a former information edge into a seven-figure cost of doing business.
The old diligence workflow next to the new one
label: The shift
left: Manual diligence
right: Agent-orchestrated diligence
caption: The work does not disappear. It changes hands, and the analyst's time moves to the end of the chain.
Sample the corpus: latest 10-K, four recent transcripts, two broker notes :: Read the corpus: a decade of filings, forty transcripts, peers, suppliers, customers
Coverage is a function of which names you had time for :: Coverage is a function of which names you told the agent to run
Reconcile the deck against the filing by hand :: Agent reconciles and flags where the two disagree
Four to six weeks to a first defensible view :: Hours to a first defensible view, then weeks of judgment on top
Analyst spends most of the time sourcing and processing :: Analyst spends most of the time deciding what the evidence means
Red flags found late, if at all :: Forensic checks run on every name by default
Where the bottleneck actually moved
Volume was never the whole problem. Somebody still has to reconcile the numbers, because the figure in the investor deck does not always match the figure in the 10-K, and the segment definition changed two years ago without a footnote anyone read. That reconciliation is where analyst hours quietly go.
Brett Caughran, formerly of Maverick, Citadel, D.E. Shaw and Two Sigma, estimated that 60% to 70% of a fundamental analyst's time goes to sourcing and processing information rather than deciding what it means. That is the share AI actually attacks. Once an analyst knows how to brief an agent, the agent does the searching, the reading, the extraction and the monitoring, and it does not sleep. Many analysts are only now working out that they cannot outwork one.
So the bottleneck moved. It is no longer finding the document. It is proving which piece of evidence is load-bearing for the thesis, and being able to show the citation when the PM pushes back.
Why agentic collection beats a traditional scraper
Plenty of tools collect documents at scale. That is not a differentiator in 2026, and it is the first thing to check when comparing AI equity research tools. AlphaSense, which absorbed Tegus in 2024, Bloomberg, FactSet and S&P Capital IQ all index enormous libraries, and expert networks like Third Bridge and GLG sell the calls that fill the gaps at $500 to $1,500 an hour.
What changed is how collection happens outside those libraries. A traditional web scraper is code written against a fixed list of sites, and it breaks the moment one of them changes its layout, which means an engineer maintains it while an analyst waits.
An agent starts from what you asked for rather than from a list of URLs. It decides where to look, rewrites its own extraction when a source moves, pulls what is relevant, reconciles it against the filings, and reports back. Matterfact gives analysts an agent that writes the scraper, runs the collection, and stands up a live dashboard on the result. We walked through how one of those gets built from a single prompt in an earlier post.
Running diligence as a chain of playbooks
Staring at a blank prompt box stops a lot of analysts who know what they want but not how to ask for it. That is a prompt engineering problem, and it is the single biggest reason AI for investment research disappoints on first contact. We wrote separately about how to write prompts for investment research, and then we pre-engineered the prompts so nobody has to.
Deep diligence on a single name runs as a chain, and every link in it is a prebuilt playbook.
variant: chain
label: The diligence chain
numbered: true
caption: Ten Initiating Coverage playbooks and nine Quality and Forensics playbooks, run in sequence on one name.
Initiation: First Read :: Orient on the business, the model and the setup
Initiation: Deep Dive :: Full primer once the name survives the first read
Management Track Record :: What this team promised three years ago against what arrived
Revenue Quality and Durability :: Whether the growth is earned or bought
Margin Architecture :: Where the operating leverage actually sits
Quality and Forensics :: Channel stuffing, GAAP versus adjusted, off-balance-sheet obligations, footnote anomalies
That chain, or something close to it, is what a good analyst runs on a new idea. Running it with agents means it also runs on the watchlist, and on the two names that have not been properly re-underwritten since initiation. There are 217 playbooks across twelve categories, so coverage stops being a function of which names you had time for.
faq heading: Frequently asked questions eyebrow: FAQ
What is the hedge fund due diligence workflow? It is the sequence between finding a name and pitching it: filing history, model build, earnings call transcripts, sell-side research, alternative data tests, expert calls, then the memo. At a discretionary long/short fund it usually takes four to six weeks per name and produces one thing that matters, a defensible view of what consensus has wrong.
How has AI changed equity research due diligence? It moved the bottleneck. Finding and reading the corpus used to be the constraint, so analysts sampled: the latest 10-K, a few recent transcripts, two broker notes. Agents read the whole corpus, including peers, suppliers and customers, which means the scarce resource is now judgment about which evidence is load-bearing rather than time to read.
How many hedge funds use generative AI? AIMA's September 2025 report, covering 150 managers with about $788 billion in AUM, put generative AI use at 95%, up from 86% in 2023, with 58% expecting to increase front-office use within a year. Adoption is no longer the differentiator. How deeply agents are wired into the research workflow is.
Can AI replace a hedge fund analyst? No. Agents do the sourcing, reading, extraction and reconciliation, which is roughly 60% to 70% of a fundamental analyst's time by Brett Caughran's estimate. The judgment about which thesis matters and how hard to push it is still the analyst's, and that is the part that gets paid.
How is an AI research agent different from AlphaSense or Bloomberg? Those platforms are libraries with search over them, and they are very good at that. An agent starts from your question rather than a query box: it decides which sources to consult, extracts and reconciles across them, flags where the investor deck disagrees with the filing, and can stand up a live dashboard that keeps updating. Search returns documents. An agent returns an answer with the citations attached.
How long does due diligence take with AI agents? The first defensible read on a new name compresses from weeks to hours, which is consistent with what Balyasny reported in OpenAI's case study on its internal research platform. The judgment layer on top still takes time. What changes is that the weeks go into thinking rather than gathering.
Read the pillar: [The Five Functions of a Long/Short Analyst Have Changed Thanks to AI](/blog/five-jobs-long-short-analyst). Related: [Pitching the PM](/blog/pitching-the-pm), [the analyst playbook for AI](/blog/analyst-ai-playbook-hedge-fund), and [inside Playbooks](/blog/playbooks-tutorial-research-workflows). Idea generation and monitoring are coming next in this series.