Will AI replace investment bankers? What the record shows in 2026

If you are a first-year analyst, your week is decks, comps and earnings transcripts. Those are exactly the tasks every AI vendor demos, so the question on the desk is no longer abstract.

The usual reading is that the analyst class goes first and the pyramid flattens. The record supports half of that. It shows the tasks leaving. It does not yet show the output arriving, and it shows AI making the judgment work worse.

Start with this week. On Bank of America’s Q3 2026 call, CEO Brian Moynihan said the bank has “about 130 different projects implemented” and expects their benefit “to go from the $250, $300 this year to $700 million next year” [source].

What “will AI replace investment bankers” actually asks

Whether AI will replace investment bankers is a question about which parts of the job a model can do to the standard a client pays for, and whether removing those parts removes the person or just the hours.

The two halves have different answers. That gap is the whole argument.

Will AI replace bankers at the big banks?

Not one for one, but headcount is already moving, and the banks describe it as automation rather than AI. Bank of America’s CFO Alastair Borthwick said “our headcount was down nearly 3,500 from a year ago as we automate processes and simplify work across the company” [source].

Moynihan, on the same call, gave “headcount down by 3,000, 1.5% year-over-year” [source]. The two executives were roughly 500 people apart and the call did not reconcile them. Neither said AI cut the jobs.

The same prepared remarks say the bank adds “headcount in our sales forces” and will “continue to invest in bankers, technology, and AI tools” [source]. Read together, the bank is shrinking process roles and buying client-facing ones. These are company statements on an earnings call, not measured results.

Will AI replace investment banking analysts?

Analysts are the most exposed, because their tasks are the ones banks already route to AI. JPMorgan says employees use LLM Suite to “generate client-ready presentations, analyse corporate earnings transcripts, compare financial documents, and synthesise data insights” [source].

The scale is real. By JPMorgan’s own count, LLM Suite “went from zero to 200,000 onboarded users within eight months” [source].

The payoff is not. JPMorgan’s chief analytics officer described “an hour saved here and three hours there” and warned the snips “often just shift bottlenecks” [source]. That is the bank that deployed it, telling you the hours do not convert cleanly.

Where AI makes banker work worse

The best controlled evidence on analyst-type work is the Harvard and BCG field experiment. It covered 758 consultants, not bankers, but the tasks were the same kind: research, synthesis and a recommendation [source].

Inside what the authors call the frontier, GPT-4 users were “completing 12.2% more tasks and completing them 25.1% more quickly on average” [source]. That is the deck and the comps table.

Outside it, on a managerial judgment task, AI users “were 19% less likely to produce correct solutions compared with those without AI” [source]. The Crimson reported the same split ([source]).

That second number is the investment banker’s job. The pitch is the frontier task. The call on whether the deal makes sense is not, and the study says the tool degrades it.

Why saved hours do not empty desks

The Bank of Korea measured the step everyone skips. It found that “AI adoption reduces average work time by 3.8 percent, equivalent to approximately 1.5 hours per week” [source].

Then it checked output. “The relationship between time savings and actual output growth is essentially zero,” the central bank wrote [source]. Central Banking’s headline put it as time saved that “does not increase output” ([source]).

This is a Korean household survey, not a bank. It still answers the question a bank’s headcount plan rests on: an hour saved is not an hour of output until someone redesigns the work around it.

The proof

TIN checked each of these against the public record before this post argued from it.

JPMorgan’s LLM Suite rollout holds the deployment scale, and shows the bank’s productivity figures conflicting across outlets with no independent measurement behind any of them.

The Harvard and BCG jagged-frontier experiment holds the randomized result: faster and more inside the frontier, less correct outside it.

The Bank of Korea’s 2026 adoption study holds the gap between time saved and output produced.

The Bank of America figures are a single earnings call, so they sit on the radar wire as reporting, not in a case file.

What each record shows about banker jobs

The bottom line

AI will not replace investment bankers, and the comfort in that sentence is misplaced. It is replacing the analyst’s task list, which is how bankers have always been trained, while the measured evidence says it weakens the judgment the job is built on. The headcount falls first at the bottom, through automation the banks name and AI they will not credit. The transferable rule: when someone tells you AI saved hours, ask which roles were redesigned to spend them.

If the answer is none, the saving is a forecast.