The edge is what's missing
The models are ready. The binding constraint is getting a specific firm's judgment out of three to five senior heads, and that's an elicitation problem.
In plenty of knowledge-work domains the capability question is settled. Models already read documents, spot inconsistencies, and draft analysis well enough. The binding constraint on agentic automation is loading them with the judgment that makes a specific firm good: judgment that lives in three to five senior heads, resists being written down, and decays. Getting it out is an elicitation problem. It has known methods, a known price, and known failure modes. The interesting engineering sits upstream of the model.
An expert’s advantage over a smart generalist is knowing where to look: which situation-type this is, which question to ask next, when to stop asking and decide. Recognition and localization carry the skill. Reasoning inside the frame, once it’s set, is close to a commodity now.
I’ll work one example in detail: private-equity deal sourcing.
The bottleneck
A mid-market PE firm’s deal-sourcing capacity is set by senior judgment, and senior judgment is scarce by construction. Evaluating a deal properly costs senior attention, and that cost sets two numbers that define the business: how many deals the firm can evaluate per month, and the minimum deal size worth evaluating at all.
The same teaser gets a 20-minute kill from a senior partner and a six-hour “maybe” from a capable VP. The partner saw something familiar and disqualifying. The VP saw a plausible company and started building a model. That delta, multiplied across every deal in the funnel, is the core of the firm’s cost structure. Most of the rest is overhead on top of it.
The shape isn’t specific to PE. A business that sells judgment tends to carry its own version of the delta, and wherever it appears it sets the same two numbers: a throughput ceiling, how much work the senior people can evaluate, and a minimum-stakes floor, the smallest problem worth their attention. Law prices matters this way; so do underwriting and engineering review. Below the floor sits work that goes unjudged because judging it costs more than the stake justifies.
The skeleton
How much of that partner’s 20-minute kill sits in the public record? Mine the Stanford and Yale search-fund studies, IESE’s research on failed acquisitions, practitioner playbooks, and buyer forums, and a few hours of work yields a skeleton pattern library for screening sub-$5M EBITDA services businesses. Six patterns, each a triple: cues, expectancy, default action.
- The owner is the business. Customer relationships run through the owner; there’s no second management layer. Expect revenue to leave with the owner. Pass, or price the transition risk with a long earnout.
- The disguised single customer. Apparently diverse customers turn out to be subsidiaries of one parent, or concentrated in one cyclical industry. Demand customer-level revenue before LOI; stress-test losing the top two.
- The reluctant seller. Prior failed sale attempts, slow document responses. Expect wasted diligence or post-close interference.
- The numbers don’t deposit. Bank deposits fail to reconcile with reported revenue. Aggressive add-backs. A revenue spike in the final six months before sale. Default action: rebuild revenue from 24 months of bank statements, and walk if it stays unresolved.
- The deal you’ve fallen for. After a long search, confirmation bias converts structural problems into fixable ones. Disciplined buyers walk away from 80 to 90 percent of signed LOIs. The pattern to recognize is in the buyer, and it’s the hardest one to see.
- The re-trade fork. Diligence surfaces a material problem; the reflexive move is renegotiating price. Experienced investors often convert the finding into better terms while protecting the post-close relationship.
The triple is the format the naturalistic-decision-making literature keeps finding when it looks at experts, and it travels: what did you notice, what did that make you expect, what do you do about it by default. A domain with a public literature yields a skeleton like this in hours, because a profession’s shared layer is, sooner or later, written down.
The skeleton takes hours, and that speed cuts both ways. Anyone can have this list by Friday; competent competitors already carry a version of it in their heads.
A firm’s edge is precisely what’s absent from that list. The seventh pattern. The override a partner applies when all six official criteria say yes and something still smells wrong. Generic AI ships with the textbook, and the textbook is the part with no edge in it.
Getting at the seventh pattern
The obvious move fails. Ask a partner “how do you evaluate deals?” and you get the training-manual version: team, market, unit economics, the stuff already on the firm’s website. Experts cannot accurately self-report their own decision process. Zacharakis and Meyer showed this with venture capitalists: the criteria investors said they used diverged substantially from the criteria that statistically explained their actual decisions. Any capture effort that just asks experts what they do will faithfully automate the training manual and miss the edge entirely. I’d bet this is a common failure mode of “AI transformation” projects in knowledge work, and the kind that rarely gets noticed.
The working method is Gary Klein’s Critical Decision Method: incident-based expert interviews. You walk the partner through specific real deals. The one killed in 30 minutes. The pass a competitor later won big on. The one that fooled them. At each step you probe: what did you notice, what did you expect, what did you ask for next, what answer would have changed your mind. The knowledge surfaces case by case, because that’s how it’s stored.
The honest price: 30 to 40 hours per senior expert, and I don’t think it compresses. The incidents are where the judgment lives, and I haven’t found a shortcut through them.
What capture produces
Done properly, elicitation yields three layers.
The firm’s pattern library: the situation-types its best people recognize on sight, including the ones missing from the public skeleton. The cue inventory: early signals like how a seller answers a document request, or the timing of a sale process; experts flatly deny having these until you walk them through incidents and the cues fall out of their own retellings. And the acquisition map: expert evaluation is a loop (read, notice what’s missing or suspicious, request specific information, re-read, repeat until conviction or kill), and a large share of senior skill is knowing the next request, plus knowing when to stop and decide.
Public sources, meanwhile, overflow with explicit rules and thresholds: “no customer over 20 percent of revenue” shows up in playbook after playbook. The scarce layers, the ones worth the elicitation bill, are the patterns, the cues, and the overrides.
Deployment follows one principle: scale the no, slow the yes. Fast rejection is almost pure gain. A fast automated yes is exactly where pattern-matching fails: judgment domains tend to give slow, noisy feedback, the condition under which expert intuition is least trustworthy. So the system prepares, flags, and kills. Humans decide the yes.
In the PE instance, every teaser killed in an hour on captured patterns, when it used to take a week of VP time, returns senior attention to the funnel. The economics follow. Make evaluation an order of magnitude cheaper and the minimum viable deal size drops. A long tail of smaller companies, currently unservable because evaluating them costs more than the position justifies, becomes addressable. That’s market expansion stacked on top of the throughput gain.
Backtesting the judgment
There’s a validation step available here that most knowledge-work automation never gets. A firm with decent records is sitting on a labeled dataset: years of inputs, the call its experts made on each one and, where it’s known, what happened after. Replay them. Give the captured system each case as it looked on the day it arrived, and compare. In PE the dataset is concrete: teasers, CIMs, and diligence files, with the pass or pursue on each.
The agreement rate is the boring output. The disagreements are the product. A deal where the system and the partner split usually points at one of two gaps: the system was missing context the partner had (a phone call that never made it into the file), or the partner was running a pattern nobody captured. Both are elicitation leads. Take the disagreement back into the interview room, walk the incident again, and either the intake spec gains a new context requirement or the library gains its seventh pattern. Backtesting would turn capture from a one-shot interview project into a loop that finds its own holes.
Two honest limits. The deal file the firm kept is contaminated by hindsight (it contains what they learned in month three) so the replay has to be built from as-of-that-day material, which is real work. And the outcomes are lopsided: passes stay unresolved forever (you rarely learn what happened to a deal you walked away from, beyond the ones a competitor won), so the backtest scores the no far better than the yes. That happens to match the deployment principle.
Where it breaks
Three failure modes stand out, and none of them is specific to PE.
Incentives. A partner’s undocumented judgment is personal leverage inside the firm. It’s why they’re hard to fire and expensive to poach. Asking them to spend 40 hours externalizing it triggers rational resistance, and the general form is uncomfortable: the people whose judgment you most need captured are the ones with the most to lose from capture. The mitigation is to frame and pay for the work as multiplying the partner: their read applied to more deals than they could personally touch, their standard becoming the house standard. I don’t know whether that framing survives contact with an actual partner’s actual self-interest. It might. It’s the open question I’d watch first.
Decay. Judgment forms inside a market regime. A pattern learned across a decade of cheap credit partly misleads in dear credit; the cues still fire, and some of them now point the wrong way. Captured judgment is a snapshot that starts aging on capture day. The countermeasure is mechanical: annotate each pattern for regime-dependence, refresh on a schedule. So capture becomes a standing pipeline. It has no end date. I’m less sure about this one than anything else here; I have yet to see evidence on how fast the decay runs, and sustaining a refresh pipeline requires the kind of discipline firms usually reserve for things with a P&L line.
Garbage-in at scale. The economics point down-market, and information quality worsens down-market: unaudited financials, owner-kept books, revenue that exists mostly in the owner’s telling. If the captured judgment includes the distrust cues (the numbers-don’t-deposit material about deposits that fail to reconcile), the system scales the skepticism along with the analysis. If it doesn’t, scaling the analysis scales the garbage. Automation amplifies whatever verification discipline was or wasn’t captured. The elicitation has to chase the paranoia as hard as it chases the pattern recognition, and paranoia is the part experts articulate worst.
Upstream of the model
The skeleton library, the CDM protocol, the backtest loop, the scale-the-no architecture: all of it runs on model capability that shipped a while ago. The project’s actual substance is 30 to 40 hours per expert of structured disagreement with what those experts believe about their own minds, plus the pipeline to keep the output alive.
Cut the cost of evaluation by ten and the minimum-stakes floor drops; a tail of work nobody could afford to judge becomes judgeable. The firms that get there first will be the ones that treated elicitation as the engineering problem. The model was ready before they started.