The case for letting AI write the research is easy to make: faster, cheaper, more consistent. The case for what quietly disappears is harder, and it matters more.
There is an assumption taking hold across capital markets. As AI becomes capable of writing research, the obvious move is to automate as much of the writing process as possible, starting with the most repetitive layer of it. Earnings updates, estimate revisions, rating reiterations, reactions to the news: AI is already good at this kind of maintenance research, and it is only getting better. The economics are hard to argue with. Analysts spend less time producing routine content and more time with clients. Firms lower costs. Coverage expands.
I think that future is both inevitable and worth having.
But maintenance research was never the whole story, and the success of AI there is quietly being read as evidence about something much bigger. If AI can competently handle the routine layer of research, the assumption follows naturally that it can take on the rest of the writing too: the thesis pieces, the initiations, the work that is supposed to carry an original view. That is a much larger claim, and it deserves a different question.
The question is not whether AI can write research. It clearly can, and across more and more of it, it will do so well. The question is what happens to an institution when AI becomes responsible for most of its writing, including the writing that is meant to generate ideas rather than simply report on them.
Experienced analysts know something that is hard to measure and impossible to ignore. Writing is not only the act of communicating an idea. Very often it is the act of finding one.
Most analysts have started a report believing they understood a company and finished it having changed their mind. A valuation framework becomes clearer because the discipline of explaining it exposes what was weak in the original thinking. Conviction arrives only after the argument has been built, not before.
That is why the sequence matters.
For a long time the order was simple. The analyst thought. The analyst wrote. The institution learned.
Increasingly the order is different. The AI produces the draft. The analyst edits. The institution publishes.
The difference looks trivial. It is not. It changes where the thinking happens. Editing is not the same activity as creating. Reviewing an argument is not the same as constructing one from first principles. Improving an explanation that already exists asks less of you than discovering one that does not.
The natural objection is that this is exactly backwards. Take the drudgery of routine notes off an analyst's desk and you free them to do more original work, not less.
I wish that were how it worked. But the time you free up does not automatically convert into original thought, because the writing was not only drudgery. Remove the task and you do not redirect the cognition. You quietly remove the occasion for it. And once AI is fluent enough to draft the routine notes well, there is no natural stopping point that keeps it confined to them. The same fluency that handles an earnings update reaches just as easily into the thesis piece next to it.
Writing was never simply the expression of thought. It was one of the environments in which thought emerged.
An AI trained on a firm's research becomes very good at reproducing its language, its structure and its analytical style. It sounds like the firm's best analysts because it learned from everything they have written. But over time something subtle sets in. Today's research is derived from yesterday's. Tomorrow's is derived from today's. The institution grows more consistent because every new document is built on the accumulated patterns of the old ones.
Consistency compounds. New ideas enter the corpus more rarely. The firm's research starts to reflect its prior conclusions more than it generates new ones.
The result is not worse research. It may well be better research. Cleaner, more consistent, more defensible. It is simply research that becomes steadily better at preserving what the institution already knows than at discovering what it does not.
Which leaves us measuring the wrong things.
We track the obvious numbers with real precision. Speed of publication. Cost per note. Analyst productivity. All of them matter.
But there is a question we almost never put a number against. How much genuinely new intellectual capital entered the institution this quarter? How many new frameworks were built? How many theses changed the way the firm understands an industry? How many variant perceptions does the firm hold today that it did not hold six months ago?
We measure output with great care. We rarely measure the creation of ideas at all. As writing becomes automated, that omission stops being a curiosity and becomes a real exposure, because the thing we are not measuring is precisely the thing becoming scarce.
The instinct in our industry is to protect research as an asset: the published note, the rating, the model. That instinct is not wrong, but it points at the output. The harder and more valuable thing to protect is the capacity that produced it. The ability to generate an idea that did not exist before is becoming the scarce, defensible thing. Routine research will be available to everyone. Original thought will not.
And you cannot protect what you cannot see.
If original thinking becomes the scarce resource, institutions will eventually need to manage it with the same discipline they once applied to documents, models and research archives.
This is the part the productivity conversation keeps missing. If an institution wants to defend its capacity to create, to reward it, to invest in it and to keep it from quietly atrophying, it first has to be able to identify it. It has to know which work carried a genuinely new idea and which reproduced an existing one. It has to be able to trace a thesis back to the person and the moment it originated, and to distinguish its own intellectual capital from what a model reassembled out of the existing corpus.
That is what protecting intellectual property comes to mean in this era. Not only guarding the asset after it exists, but preserving a faithful record of where original thinking entered the institution, so that the act of creating it can be seen, valuedand sustained.
It is how an institution keeps the scarce thing visible enough to defend.
Institutions do not get smarter simply by accumulating what they know.
They get smarter by continuing to create what they do not.
AI will become extraordinarily good at preserving yesterday's intelligence.
The firms that lead the next decade will be the ones that make sure it never quietly replaces the process, or erases the evidence, of how tomorrow's intelligence is created.