Origination as a coverage problem: screening thousands of founder-owned companies against the thesis

Start with the business problem. Why is origination the hard part for a firm your size?

For a firm our size, the deal itself is rarely the hard part. Diligence and structuring are well-trodden. The hard part is being in front of the right company at the right moment, before it hires a banker and runs a competitive auction that compresses our returns. The businesses we like are fragmented and privately held. A single thesis, say specialty distribution in one region, might contain two or three thousand companies, most of which never appear in a database because they have no reason to market themselves.

Historically, coverage was a headcount problem. Two associates building lists by hand could seriously study a few hundred companies a year. Everything beyond that was left to chance, to a broker's email, or to whoever happened to call us first. That is an expensive way to source, because the deals that reach you through an intermediary are, by definition, the ones everyone else is also seeing. The commercial value of origination sits in the companies you find first and approach directly.

Firms have tried automated sourcing for years. Why is this viable now and not five years ago?

Two things changed. The first is that language models can now read the messy, unstructured text that describes small private companies: a dated website, a trade-association listing, local news, hiring pages, customer reviews, permit filings. Older systems relied on keyword matching and structured databases, which simply do not cover this part of the market. Semantic understanding lets us screen on what a company actually does, not on how a data vendor happened to tag it.

The second change is entity resolution. The same business shows up under slightly different names across a dozen sources, and reconciling that used to be manual work. Models are now good enough to recognize that these records describe one company and to assemble a coherent profile. Compute is also cheap enough to run this across a whole sector rather than a shortlist. None of this was practical five years ago at a price a boutique could justify. I would add that the tooling is still improving quickly, so we treat our current setup as a version, not a finished system.

Walk me through how a thesis actually becomes a list of companies.

It starts with the thesis, written in plain terms: what the business does, the size range, the ownership profile, geography, and the signals that suggest a fit. We translate that into structured screening criteria and a set of softer descriptions the model can reason over.

From there the system builds a market map. It pulls from licensed company data, public web sources, hiring pages, trade directories, and filings, then resolves those into a deduplicated list of companies in the sector. Each company gets a profile and a score against the thesis. The signals matter here. Sustained hiring can indicate growth. A dated site carrying second-generation family names can indicate an approaching succession, which for us is often the more interesting signal. The model ranks candidates and explains why each one surfaced.

Then a person takes over. An associate reviews the top of the ranked list, discards the obvious misses, and researches the genuine fits before any outreach. The AI compresses weeks of list-building into a first pass we can trust enough to act on. It does not decide who we call, and it does not contact anyone.

What data goes into this, and how do you handle confidentiality and any regulatory exposure?

We separate two kinds of data and treat them very differently. Sourcing runs almost entirely on public and licensed information: websites, filings, directories, news. That carries little confidentiality risk, though we are careful about how we license and store it.

Deal data is another matter. Once a company is in a process and we are under an NDA, its financials and any material non-public information stay inside controlled systems. We do not paste that into general-purpose tools whose terms let the vendor train on inputs, because confidential information entering a model can, in principle, resurface elsewhere, and for us that is not only a data-security issue but a securities-law and fiduciary one. Our screening environment uses retrieval against our own stores rather than sending sensitive text to a public model.

Access is controlled by deal, so people see only what they are cleared to see, and we keep a record of where the sourcing data came from. Regulation here is less prescriptive than in law or medicine, but our LP commitments and confidentiality obligations set a high bar, and we design to that bar rather than to the minimum.

How accurate is the screening, and who is accountable for what it surfaces?

I would separate two failure modes. False positives are companies the system ranks highly that turn out not to fit. Those are cheap: an associate spots them in minutes. False negatives, the good companies the model misses, are the ones that actually cost us, so we deliberately tune toward broader recall and accept more noise at the top of the funnel.

We calibrate against our own history. We feed in deals we pursued and deals we passed on, and we check whether the model would have surfaced and scored them sensibly. When it disagrees with a partner's judgment, that disagreement is the useful part, because it either catches our blind spot or exposes a flaw in the model.

Accountability does not move. The score is an input, not a decision. A deal partner owns every company we approach and every one we choose not to. In our experience the right mental model is a very well-read junior analyst who never tires: fast, broad, and occasionally confidently wrong. You would not let that person commit capital, and we do not let the system stand in for judgment.

How did the team react, and what went wrong first?

The early skepticism was reasonable, and some of it was earned. Our first version leaned too heavily on keyword matching and produced long lists padded with companies that were the wrong size or the wrong business. When a tool wastes a senior person's time twice, they stop opening it. Trust is easy to lose and slow to rebuild.

What fixed it was less clever modeling and more discipline about the workflow. We rewrote how the thesis is expressed, tightened the scoring, and, importantly, made the system explain why each company appeared. Once people could see the reasoning, they could correct it, and the corrections made it better.

The harder change was cultural. Associates used to measure their contribution in companies researched. We had to redefine the job around what the machine cannot do: judgment on the shortlist, the first human conversation, and the relationship a founder actually remembers. That was a more delicate conversation than any technical fix, and it is the part I would tell a peer to plan for first.

What has changed measurably?

The clearest change is coverage. Against a given thesis we now screen roughly ten times as many companies as we could by hand, and we build the initial market map in days rather than the better part of a quarter. That means we are reaching some founders directly and early, before a process starts, which is exactly where a firm like ours wants to be.

I would be careful about the caveats. Screening a company is not the same as closing one, and the real measure of origination is capital deployed into good businesses at fair prices, which plays out over years. What we can say honestly today is that we are looking at far more of the relevant universe, our maps are current instead of stale, and more of our conversations are proprietary rather than intermediated. Those are leading indicators. Whether they convert into better returns is something I will be able to answer with conviction only after this vintage matures. We are giving ourselves more good shots, not manufacturing certainty.

What would you say to a skeptical peer, or to a founder who hears the word AI and worries?

To a skeptical peer I would say this does not replace relationships or judgment, and anyone selling it that way is overselling. It replaces the mechanical part of coverage. The quality of the output depends entirely on the quality of the thesis and the data, and in this market the data is patchy: the best companies are often the least visible online, which is a real limitation and not a rounding error.

To a founder who tenses up at the word, I am direct. We use it to understand public information about a company and to decide who is worth a genuine conversation. We are not buying dossiers of personal information, and no algorithm is deciding to acquire anyone. When we reach out, a person has read about the business and means it. Most founders find that reassuring once it is explained plainly. The firms that will get this wrong are the ones that treat volume as the point and blast automated outreach. That erodes exactly the trust this segment runs on.

Give me a specific instance where it mattered.

Here is a representative case, with the details generalized. We were working a thesis in specialty distribution, a fragmented category with a lot of second-generation, family-owned operators. The map surfaced a company that had barely any digital presence, but the model flagged a cluster of signals: the founder's name on filings going back decades, recent hiring for a general manager rather than a specialist, and a website that had not changed in years. Read together, those pointed to a possible succession moment.

That company would not have made a hand-built shortlist, because on the surface it looked dormant. An associate researched it, we made a direct and unhurried approach, and we were in a real conversation well before anything resembling a process. Whether it becomes a closed deal is a separate question with its own path. The point is that the signal combination was subtle enough that a person scanning quickly would likely have skipped it, and the system did not.

What does this mean for how a firm like yours competes over the next few years?

The origination edge in this market is shifting. It used to rest almost entirely on relationships and a physical presence in a region, and those still matter enormously. What is changing is that the coverage layer underneath them is becoming systematic. A boutique can now study a sector with a thoroughness that used to require a much larger team, which lets smaller firms compete for proprietary flow they previously could not reach.

The obvious risk is that these tools spread and the advantage compresses. If everyone can build the same map, the map is no longer the edge. So I expect the durable advantage to move back to the things that are genuinely hard to copy: the sharpness of your thesis, the conviction to act on an unpolished company, and the trust a founder places in you over a competitor. The technology raises the floor on coverage. It does not decide which companies are worth owning, or persuade a founder to sell to you rather than the firm down the street. That remains our work.