The AI conversation in M&A has spent two years oscillating between two unhelpful extremes. On one side, partners who insist AI cannot do any meaningful work in finance because it hallucinates and lacks judgment. On the other, founders selling “AI will replace junior bankers” timelines that anyone who has run a deal knows are not credible. Both positions are easy to take if you have not actually sat with the work.
The honest assessment is more interesting and more useful. There are specific parts of M&A work where AI is now meaningfully better than humans, and specific parts where it is not close. Knowing which is which is the difference between deploying it effectively and either over- or under-using it.
What AI cannot do in M&A
Judgment calls on equity story emphasis. The decision about whether to lead the IM with growth narrative or profitability narrative is not a documentation problem. It is a positioning problem that depends on reading the buyer universe, the founder’s risk tolerance, the current market sentiment in the sector, and a dozen other inputs that aren’t legible in the data room. The senior associate who has run twenty deals in industrial services has a felt sense for which framing will get the right buyers to lean in. AI does not have that felt sense. It can articulate options, it can pressure-test them, but it cannot replace the partner-level instinct about which story to tell.
Reading the room with a founder. Most sell-side mandates involve managing a founder through the most stressful financial decision of their life. The founder may be selling because they are tired, because they are sick, because they want to fund a divorce, because their co-founder pushed them out, and the advisor’s job is to figure out the actual motivation and structure the process around it. The founder will not tell you any of this in the data room. They will tell you over coffee, or after a third glass of wine at the kickoff dinner, or never. AI cannot run that conversation.
Negotiating fee structures. The conversation about whether the mandate is 1.5% with a $500K minimum or 2% with a smaller minimum is not a quantitative optimization problem. It is a relationship and signaling problem. The fee structure communicates how much the advisor values the deal, how committed they are to closing, and what their leverage looks like with the buyer universe. The partner has to know when to hold the line and when to flex, and that knowledge comes from reading the founder, the comparable deals in the market, and the advisor’s own book of work. AI can model fee outcomes. It cannot read the room.
Knowing when a buyer is bluffing. Mid-process, a private equity buyer pushes back on valuation. They cite weak comps, fading market sentiment, dilution risk from the management equity rollover. The question for the advisor is: is the buyer actually going to walk, or are they testing how firm the seller is? The answer depends on knowing the buyer’s fund dynamics, their deal pipeline, their behavior in past processes, and the personality of the principal across the table. This is exactly the kind of pattern recognition that humans with deep experience are uniquely good at and AI has no real path to.
What AI does better than humans
Cross-document consistency at scale. Verifying that a given figure appears identically across sixteen documents is a task humans are bad at and machines are perfect at. The reason analysts spend so many hours on reconciliation work is not that the work requires intelligence. It is that the human brain cannot reliably hold sixteen documents in working memory and check them against each other without errors. AI tools with proper extraction architecture do this in seconds, every time, with zero fatigue penalty.
24/7 availability and zero fatigue at hour 14. This is sometimes dismissed as a soft advantage, but it is actually structural. A senior associate at hour 14 of a 16-hour day is going to make different decisions than the same person at hour 4. The error rate goes up, the patience goes down, the willingness to chase down one more inconsistency drops. An AI tool has no fatigue gradient. The output at 3am is the same quality as the output at 9am. That alone changes what kind of deals can be done at what tempo.
Perfect recall of the data room. A 400-document data room contains more facts than a human can hold in their head at any one time. The analyst’s job is to develop a working model of the company and then validate that model against the documents. The AI’s job is to keep all 400 documents available simultaneously and answer any question against the full set in seconds. The combination, human pattern recognition layered on top of AI recall, is materially better than either alone.
How to deploy this in practice
The implication for a boutique partner thinking about workflow is straightforward. The work that AI does better than humans should move to AI immediately. The work that AI cannot do should stay with the senior team, with their time freed up by the AI handling the mechanical layer.
In practice this means: extraction, reconciliation, cross-document validation, formatting, and first-draft generation move to AI. Equity story positioning, founder management, fee negotiations, buyer dynamics, and judgment calls about deal trajectory stay with the senior team. The analyst’s role shifts from “do the mechanics under senior supervision” to “validate the AI output and develop deal judgment alongside the senior team.”
The firms that get this right will run more deals per partner, with better turnaround time, and with senior teams that are less burned out because they are spending their energy on the work that actually requires them. The firms that get it wrong will either over-deploy AI into areas where it isn’t ready, or under-deploy it and continue paying analyst rates for work that should not require analyst hours.
Both kinds of mistakes are visible from the outside. Founders know which firms are operating with what tooling. They are starting to factor it into the mandate decision. The next 18 months will sort out which boutiques figured out the line and which ones didn’t.