Will AI Replace Equity Research Analysts? The Evidence

No, and the evidence is more interesting than the question. Two studies of real analyst output published within eight weeks of each other reach opposite conclusions about whether AI-assisted research is more accurate. What neither disputes is that AI has already absorbed most of the assembly work. The part of the job that survives is judgment, and it is now the whole value.

That is a more useful answer than the two you usually get, which are "AI will replace analysts within five years" and "AI is just a tool, relax." Both are guesses. There is now real data on what happened to published research after analysts got AI, and it is worth reading carefully, because the two best studies do not agree.

What does the evidence actually say about AI and equity research?

Two working papers, both looking at live sell-side output rather than a laboratory task, reach different conclusions on the one measure that matters most.

Read the third column and the fourth column together. Both papers agree that AI-assisted reports get broader: more sources, more topics, more valuation methods. They disagree entirely on whether the forecasts inside those broader reports get better or worse.

Why do the two studies reach opposite conclusions?

Because they are measuring different things about adoption, and both explanations are plausible.

The FactSet study identifies off a platform launch, so it captures what happens when a capable AI tool arrives for everyone at once, including analysts who did not ask for it and have no process for using it. Its authors offer a specific mechanism for the accuracy drop, and it is not the one people expect: AI-assisted reports convey a more balanced mix of positive and negative information, and a more balanced mix is harder to synthesize into a single point estimate. The effect is strongest among analysts already carrying heavy cognitive demands. More information made the report better and the number worse.

The second study identifies off analyst workload rather than a tool launch, which means it is measuring something closer to deliberate use under pressure. There the gain lands hardest exactly where you would expect a labor-saving tool to help: analysts covering too many names, and analysts with weaker baseline skill. It also found that forecasts from AI-assisted reports produced stronger capital market reactions, meaning investors treated them as more informative.

The reconciliation that fits both results is unglamorous. AI reliably expands what an analyst can cover. Whether that expansion improves the conclusion depends on whether the analyst has a process for turning more information into a decision. Handed to someone drowning in names, it is a lifeline. Handed to someone already at the limit of what they can synthesize, it is more to synthesize.

What is AI genuinely good at in equity research?

The split is sharp, and it maps almost exactly onto whether a task has a checkable source.

That last row deserves emphasis because it catches experienced people. Ask a general assistant what a strategy returned since 2010 and you will get a specific, confident, fabricated number, because nothing behind the answer holds market data. The prose is not the problem. The prose is excellent. The figures inside the prose have no provenance.

The peer set row is the one that quietly matters most. Choosing which five companies a business should be compared against is not a lookup, it is the argument, and a model asked to produce comparables will produce reasonable-looking ones without ever having decided anything.

Which parts of the equity research job are actually shrinking?

The assembly layer. Pulling a decade of financials into a consistent format, normalizing across restatements, reading a transcript for the four sentences that changed, building the first draft of a company profile: this work used to occupy the junior end of a research team and it compresses well. The realistic read across sell-side and buy-side is not that analysts disappear but that firms need fewer people to produce the same coverage, which shows up first as smaller junior classes rather than as layoffs of senior staff.

What does not shrink is the part that carries reputational risk. Someone still has to decide which facts matter, what the market has already priced in, and whether the thesis survives contact with a bad quarter. That has never been the time-consuming part of the job. It is now most of the job.

There is a second-order effect worth naming. When the assembly layer is cheap, the volume of research goes up, and the scarce resource shifts from producing analysis to filtering it. A lot of the AI tooling now sold to research teams is really answering that filtering problem, and the same shift is happening to internal documents generally: once every team can generate more material than anyone can read, the binding constraint becomes finding the one answer already buried somewhere in your firm's own files, not producing another document.

What can AI not do in equity research?

Three things, and none of them looks likely to change soon.

It cannot know what happens next. No amount of reading produces knowledge of a future price, and a model that sounds certain about next quarter is expressing fluency, not information. This is the same limit that makes AI trading predictions unreliable, and it does not weaken as models improve at language.

It cannot take responsibility. A published recommendation has a name on it, and the name is accountable to clients, to compliance, and to whoever reads it in two years. Handing the drafting to a model does not transfer any part of that.

It cannot tell you whether the idea in the report ever worked. This is the gap that surprises people, because it sounds like the easiest of the three. A research platform retrieves and synthesizes what has been written. Asking whether a stated claim held up historically is a different operation entirely: it needs price history, an explicit rule, and realistic costs charged. Almost nothing in the standard research stack does it, which is why we built AI equity research testing as a separate station rather than another summarizer.

How do equity research analysts use AI day to day?

In practice, as a stack rather than a single product. A document platform such as AlphaSense or Hebbia handles search across filings, transcripts and broker research. The data terminal stays where it was. A general assistant does explanation and drafting. Models stay in spreadsheets, because that is where they are audited.

The workflow that seems to hold up looks like this. Use AI for the reading. Force the output into a claim you could state in one sentence. Check the figures against a source that actually holds the data. Then, before the claim goes into a memo, test the part of it that can be tested against history.

That last step is the one almost nobody runs, and it is cheap. If your thesis is "quality compounders outperform after a drawdown," that is a rule, and a rule can be run against twenty years of adjusted daily prices with costs charged and compared against simply holding the index over identical dates. The result is frequently uncomfortable. A large share of theses that read beautifully in a report have never beaten buy and hold, and finding that out before publication is considerably better than after.

Should you still go into equity research?

Yes, with a clear view of what the job is becoming. The path that used to work, being fast and accurate at assembly and getting promoted into judgment, has had its first rung shortened. The people who do well from here are the ones who get to judgment early: forming a view, defending it against evidence, and being willing to publish a conclusion that disagrees with the consensus and with their own prior.

The skill that appreciates fastest is verification. In a world where a competent draft costs nothing, the person who can say precisely which claims in it are checkable and which are not becomes more valuable, not less. That is an unfashionable skill and it is the one the two studies above are circling: more information helped the analysts who had a process for testing it and hurt the ones who did not.

The check almost nobody runs

Both papers describe research that got broader. Neither describes research that got more tested, because testing is not what any of these tools do. The tools read, summarize, compare and draft. The question of whether the conclusion has ever held is left to the reader, and usually the reader does not ask it.

It takes minutes to ask. State the testable claim inside your thesis in one sentence, confirm the rule the bench extracted so you know exactly what will run, and watch it execute across 20+ years of split- and dividend-adjusted daily data with 0.1% charged per trade. The verdict is HELD UP, MIXED or UNDERPERFORMED against buy and hold over identical dates, and UNDERPERFORMED comes back often. Before you rely on any number in a generated report, it is also worth reading whether backtesting is accurate and how long you should backtest, because a test run badly is worse than no test at all. The full walkthrough of the research loop is in how to analyze a stock, and the tooling comparison, including what the enterprise platforms cost and what they leave out, is on AI equity research tools and platforms.

AI is not going to replace the analyst. It has already replaced the analyst's afternoon. What it hands back is time, and the honest use of that time is checking the things nobody used to have time to check.