Like everyone else in my field, I had my moment of disbelief in front of a spreadsheet that built itself. I gave Claude a model I had assembled by hand over weeks, and it rewrote it in minutes: caught a circular reference I had stopped seeing, restructured the flow assumptions, and produced a cleaner, faster, and better-integrated version. The amazement was real. So is the question that every working financier is now asking: is AI the end of financial analysis? Must we adopt it? And if we adopt it, are we adopting our own replacement?

This essay is my attempt at an honest answer. Not a eulogy, not a hype piece. A grounded look at what the analyst’s job actually consists of, which parts of it can be mechanized, what that means for the people who do it, and what a serious professional should do about it.

The core argument, stated plainly: AI ends financial analysis as production and concentrates its value in judgment — and that boundary is not fixed. The first question is not whether the profession will be replaced, but which layer of it you are building your career on.

What AI actually automates

To reason clearly about the future, it helps to decompose the analyst’s work into concrete tasks instead of treating the profession as one undifferentiated job. Here is a rough map of where the work actually sits, and how exposed each type of task is:

  • Data gathering and cleaning. Pulling statements, market data, transaction comparables, normalizing definitions. Nearly fully automatable today, and improving quickly.
  • Model construction. Building the three statements, the DCF, the comparable analysis — converting a structure into computed cells. Largely automatable, given a clear specification.
  • Sensitivity and scenario computation. Running assumption sweeps, stress cases, and what-ifs. Fully automatable; this is computation, not analysis.
  • Report, memo, and deck drafting. Producing the structured document that presents the work. Largely automatable, with editing and verification by a human.
  • Choosing which questions matter. Deciding what the stakeholder actually needs to know, and what a fair test of the deal looks like. Requires judgment.
  • Judgment under uncertainty. Committing to an interpretation when the data is ambiguous, and being able to defend it. Requires judgment, and accountability for the outcome.
  • Verification. Testing whether the model is right — whether assumptions hold, inputs are real, outputs are consistent with what the market actually does. Requires an independent, skeptical mind; the machine checks itself poorly.
  • Communication with a decision-maker. Translating findings into a usable decision, with context and caveats. Requires trust, which is built between people.

This list explains both the panic and its mistake. AI attacks the first four items hard, and those four items are where most junior analysts — and not a few seniors — actually spend their working days. But the last four are not automation territory. A machine can assist with them; only a human can decide. The claim that AI will end financial analysis confuses the production layer with the judgment layer. What ends is analysis as production. What remains, and gets more valuable, is analysis as judgment.

And the numbers agree with this map. Goldman Sachs estimated in 2023 that generative AI could automate about a quarter of all work tasks in the US — with business and financial operations among the most exposed occupations, at roughly 35% of their tasks — and noted that most exposed jobs are only partially automatable, with 25% to 50% of their workload replaceable. Citigroup’s 2024 review went further: around 54% of banking jobs have a high potential to be automated, the highest of any sector, with another 12% to be augmented. McKinsey put the banking opportunity at $200bn to $340bn in annual value from productivity gains, and projected that up to 70% of business activities could be automated. The precise number matters less than the direction: the exposure concentrates exactly where the production layer lives.

The confusion is understandable: for decades, judgment could only be demonstrated through production. You could not prove you had judgment without building the model, running the comparables, writing the memo. The production was the proof of work. AI removes that proof-of-work loop, so we mistake losing the labor for losing the craft. But the model is not the judgment. The model is the argument written down; the judgment is where the argument starts and what it must survive.

What the history suggests

The pattern is not new, and it is worth staying calm about — while staying honest.

  • Spreadsheets. When they arrived, they eliminated the hours of hand-bookkeeping and paper modeling that preceded them. They also created the modern financial analyst, made analysis cheaper and deeper, and shifted headcount away from clerical computation toward interpretation and decision support.
  • Business intelligence tools. They automated reporting. Report production shrank; decision support grew. The tools did not remove the role — they moved the role up one floor, and they raised the floor of what “doing the job” means.

In each wave, the tool destroyed tasks, not the profession, and it changed where juniors learn. When spreadsheets arrived, juniors stopped learning the craft by grinding through manual statements and started learning by using the tool and adding judgment on top of it. Fewer juniors were needed to produce the same output.

So the honest version of this prediction is not “nobody loses anything.” It is this: the number of pure production roles will compress, and the entry requirement will shift from can you execute to can you judge. One analyst with capable AI tools produces, per unit of time, the output that several analysts produced a decade ago. That is not a catastrophe and not a salvation. It is a re-pricing of labor, and production labor was always the cheapest thing to automate.

And I should be honest about how much work this analogy is doing. The spreadsheet and business-intelligence waves automated arithmetic and reporting; this wave automates tasks that resemble reasoning, including some we currently count as judgment. Two historical cases do not prove the future — they are evidence about how labor reallocates, not proof it will happen the same way this time. The honest claim is not “history says we are safe.” It is: the task-substitution pattern is real, the production-judgment boundary was never fixed, and it is moving again, faster than in any previous wave. Whether it moves past the judgment layer is the defining open question of this decade — and anyone who writes with certainty on either side is selling something.

Does my FMVA become useless?

This is the question that keeps me up some nights, the one that follows me around the office. I passed my FMVA — the Financial Modeling & Valuation Analyst certification — and less than two months later, I was confronted with this change. The timing still stings. I had aimed my whole preparation at exactly what I loved most: Excel. The sacred trinity of statements. The discipline of the grid. I aimed a career instrument at the thing I believed the job was made of, and the ground moved beneath it while I was still holding it.

How do I describe what Excel was to me? It was my life, and my dearest companion. I built my models cell by cell; nothing was given, every formula was earned. The sheet answered only to me, and I believed, with the full confidence of someone young and proud, that command of the spreadsheet was the whole game. Then there is the memory I will never lose — the one you already know if you read my earlier story. The exam. The model collapsing mid-build, the cells looking back empty, the minutes draining — and the fix that was so stupid it almost hurt. The circularity was off. My statements were not talking to each other: the very mechanism that lets a model feed back into itself had been switched off, and I could not see it because I was standing inside it.

Two months later, Claude opened that same kind of sheet, found the very circularity I had stopped seeing, fixed it, rebuilt the model, narrating its reasoning as it went. In minutes. And I had to realize, and accept, one hard thing: Claude was a better companion to Excel than I would ever be. Not a better analyst, you understand. But in the one arena where I had invested my pride — command of the spreadsheet itself — it was superior the day it touched the grid. My dearest companion was no longer exclusively mine.

So: does the FMVA become useless? No. It becomes insufficient, which is a different and more interesting problem. The certification proves that I understand the logic — the structure, the discipline, the way the statements ought to hold together. And that is precisely what it takes to notice when a machine drifts: to catch its circularities, to feel when a number is merely confident instead of right. What the FMVA no longer proves is that I can produce an analysis faster than anyone. I cannot, and neither can you. The value of the credential has shifted from production to judgment — from “I can build this” to “I can tell whether this is true.” In an age when any model can be generated in minutes, the ability to doubt one is the only rare skill left.

Are we obliged to use it?

Yes — not because a law or a regulator says so, but because the alternative is economically untenable.

The reasoning is simple and a little cold. Financial analysis is a competitive service. Whoever produces it faster, on better coverage, and at lower cost defines the market’s benchmark. Clients and internal stakeholders do not pay for effort; they pay for the quality of the decision, at the price of the decision. If one analyst can test twice the scenarios, stress the assumptions from both sides, and deliver the same answer at half the cost, the non-augmented analysis simply loses the work. That is how tools become obligatory in every profession: not by decree, but by margin.

I have my own, small version of this. A market study I produced recently — segments, scenarios, comparables — would have taken close to two weeks by hand, and would have been thinner and more error-prone. Refusing the tool would not have preserved the old way of working; it would only have ensured that someone else got the project. That discomfort is the real obligation. It is not that anyone will oblige you to adopt the tools. It is that the market will price them into the work, and the choice left is the one between being a user and being a commodity.

The obligation is real, but its timeline is slower than either the panic or the hype suggests — and that gap is the opportunity. Adoption in finance is deliberately cautious. The CFA Institute’s 2024 employer survey found that 82% of employers say the absence of industry-wide standards is holding back adoption, and 60% describe their workforce as anxious about these tools. Citigroup’s own researchers characterized current AI use in financial services as “widespread, shallow, and inconsequential,” with most banks still at the proof-of-concept stage. The market will price the tool into the work — that is not in doubt. It is doing so over a decade, not a quarter. There is time, but only for people who spend it building the skills the machine cannot yet safely replace: verification, framing, and the discipline of the audit.

Should we let it replace us?

“Let it replace you” is the wrong frame, because it leads to the wrong decision. Replacement is not a yes-or-no; it is a question of which layer of your work the machine takes over.

Replace the production layer: yes. Data cleaning, model assembly, scenario computation, and report drafting should go to the machine. They are cheaper, faster, and more consistent there. Every hour an analyst frees from production and moves onto judgment is an hour that justifies the seat.

Do not replace the judgment layer. And by judgment I mean four specific, defensible things, not a mood:

  • Independent verification. Generative models carry a real risk of hallucination and confident error. Someone must test the conclusions against reality: do the numbers survive adversarial assumptions? Where did each input come from? Does the output agree with what the market is actually observing? This is not a philosophical preference; it is due diligence.
  • Judgment under ambiguity. An AI optimizes toward the answer implied by its training data and by the framing of your prompt. Most of real finance is ambiguous — the data points in several directions at once. Somebody has to decide which interpretation to bet on, and to own the consequence.
  • Accountability. A wrong model recommendation does not hurt the model; it hurts the people and the capital that relied on it. Responsibility is the one thing that cannot be automated, because responsibility exists only when there is a person who can lose.
  • Context and trust. Numbers are not self-explanatory to the decision-maker. Part of the analyst’s value is making the analysis usable, and trust accumulates between people, not between a person and an interface.

The real risk deserves to be stated precisely. If you outsource the judgment layer too — if you delegate the question, the verification, and the conclusion to the machine — you become the weakest possible version of what you worried about: a human middleman who adds latency without adding value. That is the version of “being replaced” that merits fear. Not the machine taking your job; you giving your job to the machine, while the margins go to the people who operate it.

And here I must catch myself before I become the salesman I just accused others of being. The boundary I drew above is a description of now, not a law of nature. Consider the most striking evidence yet: the University of Chicago study that fed GPT-4 raw financial statements and asked it to predict the direction of one-year-ahead earnings. Prompted with a structured, step-by-step analysis, the model predicted correctly about 60% of the time — beating human analysts, who averaged around 53%. Naively prompted, without that structure, the same model scored 49%, below a coin flip. Read that twice, because it contains the whole decade. The machine, unguided, fails at a task we call judgment. The machine, guided, outperforms the professionals hired for it. Judgment is not safe because the machine cannot do it; it is safe only where a person does what that study’s prompt did — imposes the structure, defines the check, owns the call. The line between production and judgment is not holding still; it is moving. The only durable position is to be on the side of the line where you are adding the structure, not to assume the line will keep you safe.

There is a slower risk worth naming as well: skill decay. Verification is a skill that stays sharp only with use. If you never rebuild a model from first principles, never check a number by hand, you lose the ability to tell an answer that is right from an answer that is merely confident. The first casualty of this era may not be the analyst. It may be the analyst’s ability to doubt.

How to use it to improve your work

Concrete practices — what I have started doing, and what I would recommend to anyone in the profession:

  • Use it as a destructive reader, not a yes-man. Give it your conclusion and instruct it to break it. Send in your DCF and ask for the single assumption that would most plausibly kill the deal. Ask it to red-team your market study as if it were a competitor. The machine has no ego and no fatigue; it will find the weak point you stopped seeing, because you stopped wanting to see it.
  • Ask for the anatomy of the answer, not the answer. Instead of “what is the value,” ask “what would have to be true for this valuation to hold?” What survives from a good engagement is not the number; it is the set of conditions under which the number is credible, and that set is precisely what a decision-maker needs.
  • Let it expand scenarios; you pick the scenario space. The tool is excellent at sweeping combinations of assumptions — thousands of sensitivity runs, stress cases, reversals of each key driver. You decide which dimensions matter; it fills the space; you interpret the results. The machine’s output tells you where the model is fragile; that is information, not opinion.
  • Keep an audit discipline on anything material. For every significant figure that moves a decision: source of the input, key assumption, and a human signature meaning “checked and believed.” If you cannot trace it, do not ship it. This rule becomes more valuable, not less, as generation becomes cheap — because the cost of producing an unverifiable number collapses.
  • Keep your hands on the work sometimes. Regularly rebuild or hand-verify a critical model from scratch. Not for the output — the machine will do it faster — but for calibration. It preserves the intuition that lets you notice when a machine has drifted, and it defends you against the skill decay described above.
  • Use it to learn faster. It compresses the ramp: new techniques, new frameworks, the strongest opposing argument to your position, drilling on unfamiliar structures. Used deliberately as a teacher, it shortens the distance from junior to competent. The side effect is that the bar for “competent” rises — which is exactly what it should do.

A sober conclusion

To answer the four questions directly.

Is AI the end of financial analysis? No. It ends financial analysis as production, and it re-prices the hours spent producing. The layer that survives — and concentrates in value — is judgment: verification, interpretation, accountability, and trust.

Will we be obliged to use it? Effectively, yes — not by decree, but by margin. The market will price the tool into the work. Refusing it will push you to the slow, expensive end of the market, and that end is where work goes to die.

Should we let it replace us? Replace the mechanical layer, yes. Replace the judgment layer only at your own expense — doing so means providing the machine’s labor without its economics, and adding no value above it.

How can it improve our work? It makes us faster, broader, and harder to fool — provided we keep the discipline of the skeptic: verify, decide, and own the conclusion.

The uncomfortable, bearable truth is that this is not a threat to the profession; it is a threat to the parts of it that were already routine, and a lift to the parts that were always the point. My advice to myself, and to anyone entering this profession now: get very good with the tools, and get even better at doubting them.

Descartes built an entire philosophy on the one thing he could not bring himself to doubt — his own thinking. Cogito, ergo sum: I think, therefore I am. Ours is a smaller, blunter, and bolder creed, and I offer it to every financier reading this: dubito, ergo sum — I doubt, therefore I exist. In the age of the machine that never tires, never flinches, and never stops generating, doubt is the one act that still proves an analyst is an analyst: more than a relay between the model and the decision. The analyst who stops doubting has already been absorbed — the machine does not need a messenger that merely repeats it. But the analyst who doubts, who verifies, who challenges, who refuses the confident number until it has earned its place — that analyst exists. And that existence is the only thing the market cannot buy for the price of a subscription.

Sources

Goldman Sachs Global Macro Research, “The Potentially Large Effects of Artificial Intelligence on Economic Growth,” March 2023.

McKinsey Global Institute, “The Economic Potential of Generative AI,” June 2023; “Capturing the Full Value of Generative AI in Banking,” December 2023.

Citigroup Global Perspectives & Solutions, “AI & Finance: Bot, Bank & Beyond,” June 2024.

A. Kim, M. Muhn, and V. Nikolaev, “Financial Statement Analysis with Large Language Models,” Becker Friedman Institute / University of Chicago, 2024.

CFA Institute, “AI in the Investment Sector: Employer Survey,” August 2024.