The thing I remember most clearly from my time embedded with M&A boutiques is the silence. Not the absence of noise; there is always some background hum of typing, the office HVAC, somebody on a call in the next room. The silence I mean is the kind that sets in around 1:30am when the analyst is two-thirds through an IM draft and has run out of anything to think about except whether the figure on page 47 of the tax memo matches the figure on page 12 of the historical P&L.
That moment, repeated several thousand times across a single mandate, is what I think most people get wrong about M&A work. The popular story is that junior analysts work long hours because the work is intellectually demanding. The actual story is that they work long hours because the work is intellectually undemanding but operationally impossible.
The 47-page reconciliation problem
The specific scene I keep coming back to is an analyst checking a single number across three documents. The number is the company’s prior-year revenue. It should appear identically in: the historical financial statements, the management accounts spreadsheet, the auditor’s report, the tax filing, and three slides of the IM. In practice, it appears with five different values across those five documents, each of which is “correct” in its own context (gross vs. net, pre- vs. post-IFRS adjustment, calendar year vs. fiscal year, with vs. without intercompany eliminations).
The analyst’s job at 2am is not to figure out which number is right. The analyst already knows which number is right. The analyst’s job is to verify that the IM uses the right one, that the source citation in the appendix points to the right document, and that no other section of the IM accidentally references a different one.
This is not intellectual work. It is coordination work. And it is the kind of work that human brains are uniquely bad at after hour ten.
The copy-paste failure mode
The other scene I see repeatedly is the copy-paste from Excel into Word. The analyst has built a clean financial table in Excel. The IM template needs the same table in Word with specific formatting. The analyst copy-pastes. The numbers come over. Half the time, the decimal precision changes. The thousands separator switches from a comma to a period because the analyst is working in a Spanish-language Excel and the firm template is set to English-language Word. A €2,400k figure in Excel renders as €2.400k in Word, which the senior associate later reads as €2.4M and revises the equity story around.
This is a real failure mode I have seen happen. The remediation cost is somewhere between a frustrated 9am meeting and a fundamentally mispriced equity story, depending on when it gets caught.
The €k versus €M unit confusion specifically is a problem that compounds across European deals. The analyst learns to spot it. The AI tool has to be taught to spot it. The senior associate has to trust that one or the other has caught it before the document goes to a buyer who will not be charitable about the error.
What the bottleneck actually is
The insight I take away from these scenes is that the bottleneck in M&A document production is not intelligence. The bottleneck is coordination across documents.
Every analyst I have watched is more than smart enough to do the work. Every senior associate is more than experienced enough to review it. What breaks is the human ability to hold sixteen documents in working memory simultaneously and verify that a change made in one ripples consistently through the other fifteen.
This is the work that gets worse as the deal progresses, not better. In the first week, the data set is small and the analyst can hold most of it in their head. By week three, the data room has grown, the financial model has gone through four revisions, and the IM is on its second draft. The probability that any given number is consistent across all sixteen documents has dropped meaningfully. The probability that the analyst can detect every inconsistency by hand at 2am has dropped further.
Why this is the right problem for AI
Cross-document consistency is genuinely well-suited to machine systems and genuinely badly suited to human ones. A well-built AI tool can hold all sixteen documents in context simultaneously, flag every inconsistency, and trace every figure back to its source. It does not get tired. It does not make copy-paste errors. It does not confuse €k with €M when fatigue sets in.
What it does poorly, and this is the part most operators outside finance miss, is judgment about which inconsistencies actually matter. Two different revenue numbers in two different documents might be the same number expressed under different accounting conventions. The AI flagging both is not enough. Someone has to know which one belongs in the IM, and why.
That is where the senior associate’s time should go. Not to spot-checking page 47 against page 12 at 2am, but to making the judgment calls about which version of the truth the document presents. The analyst’s time should go to learning that judgment by sitting in those decisions, not to mechanically copy-pasting between Excel and Word until 4am.
The reason I am building NaS_OS is not that I think analysts should be replaced. It is that I have watched too many of them spend their best working years on the coordination layer, and I think the firms that fix that layer first will be the ones whose senior associates and partners are sharper, faster, and more pleasant to work with. The analysts will still be there. They will just be doing the work that actually requires being there.