AI Will Not Replace the Analyst. It Will Replace the Grunt Work.

AI is not about to replace the equity analyst. It is about to replace the large and unglamorous share of an analyst’s week that has nothing to do with judgement: gathering filings, re-keying numbers, reconciling restated figures, and rebuilding the same model skeleton for the hundredth time.

That distinction matters, because the popular version of this debate treats “analyst” as a single job that either survives or does not. It is not a single job. It is a bundle of tasks, and the tasks sit at very different points on the value curve. Automation does not arrive for a profession all at once. It arrives task by task, starting at the bottom.

Where the hours actually go

Ask a working analyst how they spent last week and the honest answer is rarely “thinking.” It is closer to “finding, cleaning, and re-typing.”

A new company lands on the coverage list. Before a single view can be formed, someone has to pull the annual report, the quarterly results, the investor presentation, the earnings-call transcript, and often several years of history to see the trend. Then the numbers have to come out of PDFs and into a model. Then last year’s figures need checking against this year’s filing, because companies restate, reclassify, and change segment definitions more often than anyone would like. Then the model has to be built, which for most companies means recreating a structure that looks a great deal like the last one.

Only after all of that does the actual analysis begin.

The rough shape of a typical week looks something like this. The exact split varies by desk, sector, and seniority, but the pattern is consistent: the majority of time goes to preparation, and the minority goes to the part that clients are actually paying for.

Task	Nature of work	Rough share of time
Gathering filings, transcripts, presentations	Manual, repetitive	High
Extracting and re-keying numbers into models	Manual, error-prone	High
Reconciling restated and reclassified figures	Tedious, easy to miss	Moderate
Building or updating the model skeleton	Repetitive, templated	Moderate
Forming a view, writing the thesis, deciding	Judgement	Low

The uncomfortable truth of the profession is that the part everyone trained for, the judgement, is the smallest slice. Most of the calendar is consumed by work that a diligent intern could do, that no analyst enjoys, and that adds no differentiated value to a client. It simply has to be done before the valuable work can start.

The grunt work is exactly what AI is good at

Look again at the top rows of that table and you are describing the tasks that current AI systems handle well: reading long documents, pulling structured figures out of unstructured text, searching a transcript for the one comment about margins, and drafting a first pass at a summary.

None of this is the hard, contestable part of the job. Gathering is not a skill anyone brags about. Re-keying a balance sheet is not analysis; it is data entry with a finance vocabulary. Reconciling a restated figure is important, but it is mechanical: find the old number, find the new number, understand why they differ. These are precisely the places where the hours leak away, and precisely the places where software can carry the load.

The scarce resource in research was never information. It was the analyst’s attention. Everything that consumes attention without using judgement is a candidate for automation.

This is the real shift. Not a machine that decides whether a business is worth owning, but a machine that clears the runway so the human gets to the decision faster, with more of the underlying evidence already assembled and cross-checked.

The analyst becomes an editor

The emerging model is not “analyst versus AI.” It is the analyst as editor and decision-maker, supervising systems that do the gathering and the first-draft synthesis.

Think about how an editor works. They do not write every word themselves. They commission drafts, then they interrogate them: Is this claim supported? Where did this number come from? Does this conclusion follow, or is it a leap? Is anything important missing? The editor’s value is judgement applied to material that someone (or something) else assembled.

Applied to research, the workflow inverts. Instead of spending four days building a model and one day thinking, the analyst starts the week with a first-draft dossier already assembled: the filings collected, the figures extracted with their sources attached, the restatements flagged, a model skeleton populated, and a synthesis written that can be torn apart. The analyst’s week is then spent on the part that was always the point: challenging the draft, catching what the machine missed, forming a genuine view, and standing behind it.

The job does not shrink. It moves up the value curve. And the skills that matter change with it. Speed at data entry becomes worthless. Skepticism, domain judgement, and the ability to spot a wrong number that looks plausible become worth more than ever.

The new risks are the reason humans stay

If AI could do all of this cleanly, the story would end with a smaller research team. It does not, because AI introduces failure modes that did not exist before, and those failure modes are exactly why a skilled human remains necessary.

The first is hallucination. A language model can produce a figure that is wrong, formatted perfectly, and delivered with total confidence. A misread PDF, a confused unit, a number pulled from the wrong period, a “fact” that was never in any source: these do not announce themselves. They look identical to correct output. In finance, a single wrong number can invert a conclusion.

The second is over-trust. The more polished the draft, the more tempting it is to wave it through. When output looks authoritative, the human instinct to check quietly erodes. This is the dangerous part, because the value of the whole system collapses the moment the editor stops editing.

The defense against both is not better prose. It is verifiability and discipline: every extracted number traceable to its source document, every claim checkable, and a human who treats the machine’s output as a draft to be prosecuted rather than an answer to be accepted. Good tools in this space are built to make that checking fast, not to make it unnecessary. Anyone selling you an unaccountable black box for financial analysis is selling you the risk without the safeguard.

This is roughly the philosophy behind what we build at Altys Labs for Indian equities: automate the gathering and the first draft, keep the human in the loop as editor, and make the underlying evidence easy to inspect. The problem is general, and the principles are the point, not the plumbing.

The practical takeaway

The question “will AI replace analysts” is the wrong question, because it lumps a valuable job together with the low-value chores that surround it. Split them apart and the answer is clear.

• Expect the grunt work to go first. Gathering, extraction, reconciliation, and the boilerplate model skeleton are already automatable. If most of your week is spent there, that part of your week is on borrowed time.
• Move up the value curve deliberately. The durable skills are judgement, skepticism, and domain knowledge, not speed at re-keying. Invest in the part a machine cannot do.
• Treat AI output as a draft, never an answer. Its confidence is not evidence. Insist that every number traces back to a source, and check the ones that matter.
• Choose tools that make you a better editor. The right system hands you more evidence and more time to think. It does not ask you to trust it blindly.

The analyst who fears being replaced by AI is usually the one still doing the work AI is best at. The analyst who thrives is the one who hands that work over, and spends the reclaimed hours on the judgement that was always the job.