5 things a hedge fund analyst still does in May 2026 that an AI agent cannot

Ken Griffin spoke at Stanford this month and said something that landed hard inside Citadel. As Marc Rubinstein reported it on May 22, the line was that “work that we would usually do with people with Masters and PhDs in finance over the course of weeks or months [is now] being done by AI agents over the course of hours or days.” Anthropic, in the same window, ranked investment analysts as the seventh most AI-exposed profession, sitting alongside data-entry keyers and customer-service representatives. Citadel manages $68 billion in assets, has produced $90.4 billion in net gains to investors after fees, employs roughly 270 PhDs, and has booked positive returns in more than 80 percent of months since launch. Rubinstein notes the same firm has also abandoned 195 of 200 generative AI projects it started. Something is happening, and something is not. Both of those things are worth being precise about before we decide which active managers we are still happy paying.

The wider pattern is not just one CEO at one conference

The 2026 InvestOps survey from SimCorp polled 200 executives at asset managers, pension funds, and insurance companies with at least 10 billion dollars in assets under management. It found 70 percent of buy-side firms are now actively using AI in the front office, up from 10 percent a year earlier. That is a 60 percentage-point jump in 12 months, which is the steepest enterprise-adoption curve I can remember reading for any technology since cloud.

The agent tools that flow from that adoption sit on top of a remarkable corpus. AlphaSense Deep Research, the buy-side research agent that launched last year and was expanded for institutional accounts in early 2026, runs across more than 500 million high-value business and financial documents. Bloomberg’s ASKB and FactSet’s research agents do similar work on adjacent corpora. The output is what Griffin described: a company primer that took a junior analyst a week now lands in an hour. A screen of forty potential merger-and-acquisition targets that took an associate a long weekend now arrives before the partner finishes lunch.

Some funds have leaned in completely. Sydney-based Minotaur Capital employs no traditional analysts, runs the entire research process through a proprietary system called Taurient that processes 5,000 news articles a day, and returned 27.0 percent in fiscal 2025. Walleye Capital sits at the other end of the augmentation spectrum: 75 percent of its 400-person team uses ChatGPT daily, the flagship Opportunities fund returned 14.7 percent net in 2025, and the firm has built an internal tool called Current that ingests transcripts and broker reports for portfolio-manager decision support. Even the most human-channel-dependent activity, the expert-network call, is now being automated by platforms like Checkmate Research, where one analyst at a 10-billion-dollar long-short fund publicly said “I can do twice the number of expert calls.”

The mechanism is a 26-year-old SEC rule that drew the line

In August 2000 the SEC adopted Regulation FD, the rule that ended a long era of selective disclosure. The text is short. When a public company intentionally shares material non-public information with certain market professionals, it must share it simultaneously with everyone. Whisper to a favorite sell-side analyst at a closed-door lunch and the company has created a Reg FD violation. Unintentional disclosure: the company must promptly make the information public, within 24 hours or the start of the next trading day.

The structural result was that companies funneled almost everything that previously moved through private back-channels into uniform public disclosure. The 10-K. The 10-Q. The 8-K. The earnings release. The earnings call, both the prepared remarks and the analyst question-and-answer session. The investor presentation. The press release. The sustainability report. The annual proxy. That is the corpus.

Twenty-six years later, that corpus is exactly the surface AI agents now consume at a pace no human can match. The public side of the Reg FD line is fully agentic.

But Reg FD did not police the other side. A one-on-one conversation in which no material non-public information changes hands is not a Reg FD event at all. The investor relations officer can take an analyst’s call and answer questions all afternoon, as long as every answer stays inside what is already public. Management can sit across from an analyst at an industry conference and respond to her questions in carefully bounded public statements. A former employee at a covered company can return an analyst’s third call. The rule permits all of this and AI agents do not legally or structurally participate in it. The residual human analyst job has migrated to the side of the line that the rule defined as out-of-scope.

Five conversations that still belong to humans in May 2026

1. Placing the call to investor relations

The investor relations officer at a mid-cap industrial company will pick up for a human analyst at a 5-billion-dollar long-short fund. She will not pick up for an agent calling on the same fund’s behalf. The conversation is not a Reg FD event because the investor relations officer is trained to answer only with information already in public materials. What the human extracts is tone, hesitation, which of the three plausible reads of last quarter’s gross margin the officer subtly emphasizes when pressed. None of that lives in the transcript that the agent will read tomorrow.

2. Sitting across from management at a conference one-on-one

The 20-minute private meeting at the Bernstein conference or the Morgan Stanley technology-media-telecom week is the canonical Reg FD-permitted private interaction. The chief financial officer answers in carefully bounded public statements; the analyst reads everything the answer leaves out. Rubinstein noted on May 15 that Tesla’s recent earnings call had 4,900 participants submitting 86 questions, of which only 9 were answered, and only 5 Wall Street analysts were permitted to ask anything at all. The narrowing of the public call has only made the private conference one-on-one more valuable.

3. Running the channel check

A channel check is a phone call to a downstream customer or upstream supplier of the company under analysis. The question sounds like, “Are you placing the same orders you placed in February, or has something slowed?” The information is not in any data feed because nobody has collected it yet anywhere. AI-mediated expert-call platforms like Checkmate Research are advancing fast, and twice the number of calls is a real productivity gain. But twice the number of agent-conducted calls is still half the depth of the human-conducted version, because the customer on the other end answers more freely to a person they have spoken to before.

4. Maintaining the former-employee relationship

The third call to a former product manager who left the company two years ago is the one that produces signal. The first two were polite. By the third, the analyst has earned the right to ask the question that matters, which is usually about whether the launch slipping by six weeks reflects a road-map problem or a leadership problem. Expert networks like GLG and Third Bridge can match an agent to a contact in minutes. They cannot manufacture the three-year working relationship that makes the former employee return the third call and answer the question that the first call could not reach.

5. Reading the portfolio manager’s face after the thesis

The 5-minute conversation in the portfolio manager’s office, where the analyst lays out the thesis, is the moment any pitch is killed or sized up. The portfolio manager listens, asks two questions, and the analyst reads the room. Some heads of the desk trust a discounted-cash-flow framing on this kind of stock; others want a sum-of-the-parts; others want the comp set with a margin overlay. The skill of reading which framing this specific portfolio manager trusts on this specific name is tacit firm-specific knowledge that does not transfer to an agent because it does not transfer to a written document either. Two Sigma’s Matt Greenwood calls AI “the operating system for how quantitative research and investing work.” Operating systems do not read faces.

What might break this read over the next few years

Three things, honestly. The first is that Minotaur is already running with no analysts and beating the benchmark, which suggests there are corners of the equity universe where the public-corpus alpha is enough. The second is that AI-mediated expert-call platforms keep getting better. If Checkmate or its successors can credibly conduct an expert call with the depth of a senior analyst, channel-check work begins to migrate to the agentic side of the Reg FD line. The third is that the rule itself could move. Reg FD was written when analysts had landlines and a research budget. Whether one-on-one private conversation with an agent on one side still counts as a permissible interaction under the original rule is a question the SEC has not yet been asked to answer directly. The probability that it gets asked in 2027 feels meaningfully above zero.

The honest read is that the human-only column is narrower than it was three years ago and probably narrower again three years from now. The honest read is also that BCG’s 2026 asset management report found the productivity gains from AI in asset management remain trapped in legacy processes, with no measurable bottom-line impact yet. The corpus is being consumed faster; the decisions are not yet being made better.

The residual moat of the human analyst is precisely the channel that Regulation FD defined as out-of-scope, which means the rule the SEC wrote in 2000 to limit selective disclosure is also the rule that now defines where AI cannot fully replace a human.

What changes for the active managers we pay

Most of our serious money sits with managers who claim differentiated insight on individual names. In May 2026 the right question to ask each of them is no longer “do you use AI” because everyone does, and that question has stopped being informative. The question is what the manager’s footprint looks like on the Reg-FD-excluded channels. How many investor relations calls per analyst per week. How many private conference one-on-ones per name per quarter. How deep the expert-network relationships actually run, and whether they are firm-level or analyst-level. A quarterly letter that says “we spoke to” more than once per page is signaling that the channel side is still alive. A quarterly letter that reads like an output of the same agent everyone else owns is signaling something else.

1. Ask each active manager for its analyst-to-name ratio.

   Not how many analysts in total, but how many names each analyst actually carries. If the ratio has been widened by AI from 20 names per analyst to 80, the channel work has thinned even if nobody on the call is saying so.

2. Ask how many private one-on-ones each analyst books per quarter.

   Conference one-on-ones, investor relations check-ins, channel calls with named counterparts. The number tells us what the management fee is actually paying for in 2026.

3. Ask which expert-network relationships are firm-level versus analyst-level.

   Firm-level relationships survive turnover and compound. Analyst-level ones walk out the door with the analyst. The split is a real measure of institutional depth.

4. Read the quarterly letter for tacit-knowledge tells.

   Phrases like "we spoke to," "in conversation with," and "at the conference last week" suggest the channel side is still alive. Pure data-narrative letters that could have been written by any agent on any corpus suggest the manager is running on the public side only, which is the side we can run on ourselves for free.

It is possible the Reg FD threshold itself moves over the next decade as agentic tooling crosses into the private-conversation side, and the SEC is forced to decide whether an agent on the buy-side end of a one-on-one with an investor relations officer is still a Reg FD-permitted interaction. The question worth sitting with is not whether agents will eventually place the investor relations call. It is whether the rule that drew the line in 2000 will still draw it in the same place when they can.