Every investment committee has a version of the same moment. A partner stops on a single figure, perhaps an exit multiple, a churn assumption, or a discount rate, and asks the analyst why it is what it is. The question is rarely about the number itself. It is a probe. If the analyst can walk backwards from that cell through the reasoning that produced it, the partner extends trust to the rest of the memo. If the analyst hesitates, the whole document is suddenly on the table.
That exchange only works if the person defending the work is the person who did it. This is precisely the assumption that careless AI adoption quietly breaks.
The trap looks like productivity
When a model can produce a polished first draft of an IC memo in minutes, the rational-seeming move is to let it, and to spend the saved hours editing. The output improves. The pace improves. Everyone involved can point to something that looks like progress.
What has actually happened is the substitution of one cognitive act for a much weaker one. Producing an argument forces you to sit with the data, decide what matters, and commit to a position that could be wrong. Reviewing a plausible argument someone else produced demands far less, and it feels more complete than it is: a fluent draft reads as though the thinking has been done, so the reader's scepticism relaxes at exactly the moment it should sharpen. The analyst becomes an editor of a draft they cannot fully defend, and nothing on the page reveals it. Quality looks the same. The judgment underneath it thins out.
There is a difference between owning an argument and having approved one, and an investment committee is designed to find it.
Why investment work is exposed
Some professions can absorb this quietly. Investment work cannot, because its entire review structure presumes authorship. A memo passes from associate to principal to partner to committee, and at each stage the reviewer samples: they probe two or three load-bearing assumptions, watch how the author responds, and extend or withdraw trust accordingly. Sampling is efficient precisely because defending one number well is evidence that the author can defend the others.
Break the authorship assumption and the sampling breaks with it. The analyst who edited a generated draft may survive the first probe, because models are good at making assumptions sound reasoned, and fail the second, or survive both and carry an undetected error into a signed term sheet. The erosion also compounds. Analysts learn judgment by producing work and having it taken apart; the partners who take memos apart learned to do it by having their own taken apart. A firm that lets its juniors skip the production step is not saving time. It is quietly cancelling the apprenticeship that produces its future partners.
Invert the workflow
The answer is not less AI. It is a different sequence. The analyst produces the first pass, the position, the numbers, and the reasoning, before the model sees any of it. Then the model is brought in as an adversary rather than an author: to attack the thesis, hunt for disconfirming evidence, stress the assumptions the analyst is most attached to, and check the arithmetic and the sourcing. We call this AI as adversary, not author, and it is the principle behind everything we do.
The inversion matters because the model is at its most valuable, and least corrosive, in the critic's chair. As an author it displaces the formative work; as an adversary it multiplies it. An analyst whose thesis has survived serious red-teaming walks into committee with better answers than either the unaided analyst or the prompt-and-polish one, and every answer is the analyst's own.
What a firm can do about it
None of this survives on good intentions, because the prompt-and-polish route will always be faster on any given Tuesday. It has to be held in place by structure. Review norms can require that the human contribution be legible: a memo arrives with its formative work attached, and reviewers are entitled to ask which parts the analyst produced and which the model challenged. Partners can make provenance a normal question in committee, asked without ceremony, the way “why this number” is asked now. Senior-led calibration sessions surface drift in individual analysts before it compounds across the team: working a live deal problem together, unassisted first, then with the model as critic. And occasionally requiring a fully unassisted first pass does for the firm's bench what an audit does for its books: it verifies that the capability everyone assumes is there is actually still there.
Firms that get this right will be faster and sharper than the firms that merely adopted the tools, because they will have kept the one asset the tools cannot replace. The firms that get it wrong will not notice for a while; their documents will look excellent. Then one afternoon in committee a partner stops on a number and asks why, and nobody in the room owns the answer.
If that question is one your firm has started asking itself, we are glad to talk it through.