Buy-Side Research With Claude: From Earnings Calls to Thesis Tracking

The honest answer to how do hedge funds use AI and how do asset managers use AI for research is less dramatic than the headlines suggest: it's mostly compression. The reading, cross-checking, and first-draft writing that used to eat a junior analyst's week get compressed into something that takes an afternoon, with the analyst spending the saved time on the parts of the job that actually require judgment, like deciding whether a changed risk disclosure matters or whether a beat was quality or one-time.

Firms across the industry have described building AI directly into the tools analysts already use rather than standing up a separate research platform. Some have said publicly that analysts use it inside a spreadsheet to build and stress-test coverage models and pressure-test their own assumptions before a name goes to the desk. Others have used it to power early versions of an internal research assistant that writes code, builds visualizations, and works through complex financial analysis, treating it as infrastructure inside an existing analyst workflow rather than a replacement for one.

What doesn't change is who is accountable for the number. A model that's wrong because an analyst trusted a flagged reconciliation without checking the source transcript is still the analyst's model. The firms that have talked about this publicly describe it the same way: faster first passes, same sign-off discipline.

Every one of these workflows produces something that looks finished before it actually is. A reconciliation between a call and a model can misread a hedge in management's language as a confirmed number. A filing diff can flag a wording change that's genuinely immaterial, or miss a change made in a format it wasn't told to look for. None of that is a reason to avoid the tool. It's a reason to treat every output as a draft that needs a source check, not a finding that's already been verified.

The same caution applies to thesis tracking. A tool that's good at pattern-matching new disclosures against old sell conditions can also make an analyst feel more confident than they should about a thesis that's actually eroding, if the conditions it's checking against were written loosely in the first place. The fix isn't a better tool. It's writing sharper, more falsifiable sell conditions at initiation, so whatever checks them later, human or otherwise, has something concrete to check.

Initiation drafts carry the same rule in a different form. A first-pass structure can save real time, but a draft that reads as polished and complete invites a reviewer to skim rather than check. The discipline that holds is simple: every number in a draft gets traced back to its source before the draft goes anywhere, and the final view gets rewritten in the analyst's own words rather than lightly edited from the first pass.

It's worth separating two things that sound similar. A plain summary of an earnings call tells you what was said. AI earnings call analysis, in the sense that matters for coverage, tells you where what was said disagrees with what your model already assumes. The second is far more useful and far easier to get wrong, because it requires the model lines to actually be fed in accurately and the comparison to be checked against the transcript, not taken on faith.

The practical habit that keeps this honest is small: always paste in both sides, the transcript and the specific model assumptions, rather than asking for a general take on how the quarter went. A general take is a summary with extra confidence. A side-by-side reconciliation, checked against the source language, is closer to what the analyst was already going to spend an afternoon doing by hand.