The Sourcing Machine Nobody Audits: What AI Deal Origination Actually Changes

Every mid-market fund now runs some version of the same pitch internally. We will use AI to widen the funnel. We will screen more companies, faster, on more signals, and we will find the ones our competitors miss. The budget gets approved, a data vendor gets selected, a screening model gets built, and the origination team starts receiving a weekly list of targets ranked by fit score.

Almost nobody audits what happens next.

The reason this matters is structural rather than technical. Proprietary deal flow has always been the most valuable and least examined phrase in private equity. It usually means one of three things: a relationship nobody else has, a piece of information nobody else has, or a willingness to do work nobody else will do. AI origination affects all three differently, and most funds are deploying it as though it affects only the second.

What the machine actually does

Strip away the vendor language and an AI origination stack does four things. It aggregates company data from public and licensed sources. It normalises that data into a comparable structure. It scores companies against a fit definition. And it surfaces change events, a new hire, a filing, a funding round, a website change, that might indicate a company has become approachable.

Each of those four is genuinely useful. None of them is proprietary.

The aggregation layer is bought. The normalisation layer is a commodity. The change detection layer is bought from the same three or four vendors everyone else uses. Only the fit definition is yours, and in most funds the fit definition is a set of filters that any competitor could reconstruct in an afternoon: revenue band, sector codes, geography, ownership status, growth rate.

The uncomfortable arithmetic is that if your screen is reproducible, your output is reproducible. Two funds with similar theses, buying from the same data vendors, running similar filters, will surface substantially overlapping lists. The machine did not find you a company nobody knew about. It found you a company that was always in the database, and it found it for your competitors at the same time.

Three of the six layers are bought and confer no advantage. Most of the AI origination budget is spent there. The three layers that do confer advantage receive the least attention, because they cannot be procured.

The three questions that separate edge from expense

One. Is any part of your signal set actually private?

A signal is private if a competitor cannot obtain it by writing a cheque. Very few of the signals in a standard origination stack qualify. Headcount growth, job postings, technology footprint, funding history, and web traffic are all purchasable.

What is genuinely private is the information a fund generates through its own activity. Every company you diligenced and passed on, and precisely why. Every management team you met and how they presented. Every seller who was not ready in year one and gave a reason. Every bolt-on your portfolio companies evaluated. Every reference call. That corpus is unique to your firm, it grows every quarter, and almost no mid-market fund has it in a queryable form.

A fund that has diligenced two hundred companies over eight years holds two hundred structured opinions no competitor can buy. Feeding that into the screen is the difference between an AI stack that reproduces the market's view and one that encodes yours.

Two. Does the screen encode a thesis or a demographic?

Most screens describe what a target looks like. Revenue between thirty and one hundred million. EBITDA margin above fifteen percent. Founder-owned. Three specific SIC codes. That is a demographic, and demographics are reproducible.

A thesis describes why a specific kind of company is mispriced. Fragmented service categories where the largest player holds under four percent share and where regulatory change is about to raise the cost of compliance for subscale operators. That is a thesis. It is harder to express as a filter, it requires judgment to encode, and it is much harder for a competitor to reconstruct because they would first have to arrive at the same view of the world.

The practical test is simple. Show your screen definition to someone outside your firm. If they can tell you what you are looking for but not why, you have a demographic. This is the same distinction that separates a real reframe from a relabelling, discussed in the strategic reframe lever.

Three. What happens after the list is produced?

This is where most AI origination programmes quietly fail, and it has nothing to do with AI.

The model produces a ranked list of two hundred companies. The origination team has capacity to run perhaps thirty meaningful outbound conversations a quarter. The list therefore gets truncated to the top thirty by fit score, and the outbound motion is the same templated sequence the team was running before the model existed. The founder receives a message that reads like every other message she received that month.

Widening the funnel without widening the throat of the funnel does nothing. Worse, it creates the appearance of progress. The dashboard shows two hundred qualified targets identified this quarter, up from forty. The conversion to first meeting has not moved, because the constraint was never identification.

The auction problem nobody says out loud

There is a second-order effect that deserves more attention than it gets. If AI origination becomes standard across the mid-market, the population of companies that are visible, scored, and contacted expands for everyone simultaneously. The previously quiet founder-owned business in a fragmented category is now receiving approaches from eleven funds instead of two.

That founder learns quickly that she has options. She hires an adviser. What was a proprietary conversation becomes a process. The industry has, collectively, spent money to convert its own off-market opportunities into auctions.

This is not a reason to opt out. A fund that does not run modern origination while its competitors do is simply late to the same auction. But it does change what the investment is for. The honest framing is that AI origination is now table stakes for maintaining position, not a source of advantage. Advantage has migrated to the three layers the vendors do not sell.

It also raises the value of everything that happens after the target is identified. If the entry point is contested more often, the differentiator becomes the quality of the operating thesis you can articulate before the seller has chosen a buyer, which is the argument in the pre-close operating veto, and the credibility of what you did with the last three companies you bought.

What to audit, and when

If your fund is running or building an AI origination capability, four questions produce more clarity than any vendor demonstration.

What proportion of our scoring inputs could a competitor purchase tomorrow. If the answer is most of them, the model is reproducing the market rather than differentiating from it.

How many companies did we pass on in the last five years, and is that decision history available to the screen in any structured form. In most funds the answer is that it lives in individual partners' memories and in unindexed IC memos. That is a private asset being left on the floor.

Of the targets the model surfaced last quarter, how many did we contact, and how many would we have found anyway. The second half of that question is the one nobody asks, and it is the one that measures incremental value.

Has our conversion from contact to first meeting changed since deployment. If identification improved and conversion did not, the constraint was downstream and the investment was aimed at the wrong place.

The part that does not automate

The persistent truth about origination is that the hardest step has never been finding the company. It has been earning a conversation with an owner who was not looking to sell, at a moment when she might, on the strength of something she believes about you rather than about your fit score.

Machines are good at the first part and structurally incapable of the second. A fund that treats origination as a search problem will build a very efficient way to arrive in a crowded room. A fund that treats it as a trust problem, and uses the machine to buy back the hours that used to go into list-building, will spend those hours where the difference is actually made.

The technology is real and worth having. Just be precise about what you bought. You bought speed. Whether that speed becomes advantage depends entirely on what you point it at, and on whether the three layers no vendor sells you are the ones you have actually built.

VCI Institute in collaboration with Mohamad Chahine
Published 19 October 2026

About the VCI Institute

The VCI Institute is a nonprofit dedicated to building practical capability and shared standards for value creation in private equity. The Institute publishes operator-grade frameworks and runs certification programmes for operating partners, portfolio company executives, and value creation analysts. You can see what each programme actually covers before deciding. Analysis published here draws on the Institute's certification curricula and on structured review of mid-market transaction patterns rather than on any single proprietary dataset. Where a figure is directional rather than measured, it is described as such.

Further material is available in the Institute's Insights library and its free resource library of templates, checklists, and case snapshots.

© 2026 VCI Institute. All rights reserved. The frameworks, terminology, and analysis presented in this article are the intellectual property of the VCI Institute. Reproduction or derivative use without written permission is prohibited. Citation with proper attribution is welcomed.