The reading and the judgment.

Before anyone commits money, signs, or takes the position, somebody has to read: the filings, the transcripts, the data room, the notes from forty calls. It is the least visible part of the finished memo and it eats the most hours. It also happens to be where the machines actually help, which makes it worth being precise about where the help stops.

Research is really two jobs wearing one name. There is taking in what exists — finding it, compressing it, noticing what a document actually says — and there is deciding what it means. The first job is bounded by the material. The second is bounded by judgment, and it is the one an experienced team is actually paid for.

At the first job, current systems are genuinely good. A well-built one can read a data room in the time an analyst reads a cover letter, flag the indemnity clause that differs from the last three deals, and assemble a first pass with citations back to the sources. For teams whose constraint is reading hours — and at most shops, that is the constraint — this changes what a given headcount can cover.

At the second job they remain unreliable, and not in a way that is going away this quarter. Stanford’s AI Index measured hallucination this spring across twenty-six frontier models: somewhere between 22% and 94%, depending on the model. Buried deeper in the report is a stranger finding. Tell a model a false thing as your own belief and it tends to go along — one leading model’s accuracy fell from 98% to 64% on identical facts once the error arrived framed as the user’s opinion. These systems complete patterns and defer to the person in the room. That is a poor temperament for the one seat in the process where resistance is the job.

Meanwhile the industry has moved in. Mercer’s 2026 survey found 55% of asset managers with AI in at least one investment process and 91% planning to use more of it within the year; AIMA’s last study found half the managers under a billion in assets running these tools with no restrictions at all. Adoption is settled. Design mostly is not.

The design that works respects the boundary between the two jobs: compress the reading, protect the judgment. In practice that means systems grounded in a firm’s own material — its data room, its past memos, its call notes — that answer with citations and assemble rather than opine. The test of such a system is not whether its output sounds like an analyst. It is whether the team’s hours move from locating things to weighing them.

There is an upside here that gets less attention than the risk. The same capability that compresses a data room lowers the cost of curiosity. A two-person firm can now read a whole industry in a week — the filings, the trade press, the patents — for roughly the price of asking. Theses that were too expensive to research get researched. Adjacent opportunities that would once have needed an associate pool get explored on a Tuesday. For investors who also build things, that may be the more interesting consequence: not faster diligence on the deals that show up, but a wider set of ideas worth taking seriously in the first place.

The honest limits, briefly. A corpus-bound system cannot surface what nobody wrote down, and a lot of diligence lives off the page — the pause before a reference answers a simple question. And a team that stops reading entirely will eventually stop noticing. The point was never to end the reading. It is to choose where it goes.