Where AI Belongs in the Banking Workflow — And Where It Doesn't

There is a version of the AI pitch that says the machine now does the banker's job, and a version that says it does nothing a spreadsheet couldn't. Both are wrong, and a firm that believes either will deploy badly. The useful question is narrower: across the actual IB workflow, where does AI reliably carry the work, and where does it hand off to a senior banker? The line is real, and it is worth drawing precisely.

Where AI carries the work.

First drafts of structured documents. The pitch book, the IC memo, the CIM section — anything with a known structure and inputs the firm already holds — is where AI is strongest. It assembles the first pass: positioning, precedent transactions, valuation framing, risk. This is the work that consumes the most junior hours and the least judgment, and moving it is almost pure gain.

Comparable company analysis. Building the comp set, screening for the right peers, and refreshing multiples as the market moves is mechanical, high-volume, and error-prone when done by hand under time pressure. AI builds it and, more importantly, keeps it current between the day it's built and the day it's presented — the gap where stale comps quietly embarrass a team.

Model population and refresh. Rekeying a revised management case into an operating model at 9pm is not analysis; it is data entry with high stakes. AI populates the model from source documents and carries assumptions forward, so the analyst reviews the logic instead of retyping the inputs.

Institutional recall. Reconstructing a coverage relationship's history, surfacing what the IC asked on a comparable deal, or finding the firm's prior look at a target — this is where AI with real memory is not just faster but categorically better than a person searching an inbox.

Where the senior banker is still the job.

Pricing and structure judgment. What a business is actually worth, how to structure the consideration, and where the negotiation has room — these rest on pattern recognition across cycles and deals that does not reduce to the inputs on the page. AI frames the range; the banker sets the number.

Reading the room and the relationship. A coverage relationship is built on trust, timing, and a read of what a client or counterparty actually wants, often the opposite of what they say. This is relationship intelligence, and it lives with the senior banker. No system covers it, and a firm that pretends otherwise damages the franchise.

The judgment call under ambiguity. When the comps disagree with the DCF, when the management case is optimistic in a way the model cannot flag, when a deal should be walked away from despite the fees — these are the decisions the MD is paid for. AI can surface the tension. It cannot own the call.

Accountability. A memo goes to committee with a name on it. The person who signs it is answerable for it, and that accountability is not delegable to a model, however good the draft. The senior banker reviews, presses, and owns the output — which is exactly why the draft arriving fully built is a gift and not a threat.

The honest synthesis is that AI is very good at the assembly and recall that surround judgment, and not a substitute for the judgment itself. The firms that win with it will not be the ones that claim the machine replaced the banker, nor the ones that dismissed it as autocomplete. They will be the ones that put AI on the reassembly and kept the banker on the decision — and were precise about which was which.

YODA AI · RAM Core.