How Thornwick Capital made 2,000 deal documents answer questions, with citations

Start with the problem. What was actually slowing the team down before any of this?

The bottleneck was never a shortage of information. It was that our information sat in forms no one could search. Every deal generates hundreds of pages: the CIM, management decks, quality-of-earnings work, legal and insurance diligence, and our own memos explaining a decision. Once a deal closed or died, those files became a dead archive. If an analyst wanted to know how we had thought about customer concentration in a similar business two years earlier, the practical answer was to ask whoever happened to remember, or to reread a folder. On a small team, the senior partners effectively were the search index, which does not scale and creates a single point of failure. We were paying skilled people to relocate facts rather than judge them, and slow retrieval quietly lengthened every diligence cycle. The cost never showed up as a line item, so it was easy to ignore, but it was real.

Firms have discussed knowledge management for years. Why is this viable now and not five years ago?

Two things changed. Earlier semantic search could find a relevant document, but it handed you a file, not an answer, and you still did the reading. A language model on its own can write a fluent answer, but it will invent specifics when the underlying material is thin or contradictory, which is unacceptable when a number ends up in an investment memo. Retrieval-augmented generation joins the two: the system first retrieves the actual passages from our documents, then the model composes an answer constrained to that retrieved text, with a citation back to the source page. The grounding is the point. The economics also shifted. Embedding and hosting a few thousand documents is now inexpensive, and retrieval quality is good enough that the cited passage is usually the one you would have chosen yourself. Five years ago the answer looked plausible but you could not check it quickly. Now every claim carries its receipt.

Walk through how the system actually answers a question, step by step.

When we ingest a document, we parse it and split it into passages, with some overlap so a sentence that matters is not cut in half at a boundary. Each passage is converted into a numerical representation and stored in a vector index, along with metadata: which deal it belongs to, the document type, the date. A question gets converted the same way, and the system retrieves the passages closest in meaning, not merely those sharing keywords. A reranking step then orders them by genuine relevance. The model writes an answer using only those passages and cites the page it drew from. If the deal room does not contain the answer, the correct behavior is to say so rather than reach into general knowledge. In our experience the chunking work matters more than which model we use. Getting the passage boundaries and overlap right does more for answer quality than swapping in a larger model would, which surprised the team at first.

These are confidential documents, and some carry material non-public information. How do you handle access and security?

This is the part we spent the most time on. The files are confidential, and some contain material non-public information, so access cannot be a free-for-all. Retrieval is permission-aware: each passage carries the access rules of its source, and those rules are enforced at query time, so a person only retrieves from deals they are cleared to see. That lets us keep information walls intact rather than dissolving them into one searchable pool. The environment is isolated to the firm, data is encrypted, and the documents are not used to train any external model, which we confirm contractually. Every query and answer is logged, so we have an audit trail of who asked what and what was returned. We treat the system as an extension of the confidentiality obligations we already owe under our NDAs. The technology did not lower that bar. If anything, building it forced us to write down access rules we had been carrying informally in people's heads.

How do you keep it honest, and who is accountable for what it says?

The citation is the control. Every answer points to the specific page it came from, so verification is a click, not a research project. Nothing the system produces is treated as fact until a professional has read the cited source and confirmed it, and that is a firm rule before anything reaches an investment memo or the committee. We evaluated it the way you would assess a junior analyst: first, does it retrieve the right source, and second, is the answer actually supported by that source rather than a fluent gloss on it. Retrieval correctness is the first gate, faithfulness to the text is the second. Accountability sits with the person, not the tool. The analyst who cites a figure owns that figure, exactly as they would if they had found it by hand. The system is fast and it is literal, but it does not exercise judgment, and we have been careful not to let anyone treat it as though it does.

What did adoption look like inside a small, skeptical team? What went wrong first?

Skepticism was healthy, and I encouraged it. The first version was worse than people expected. We had chunked documents crudely, so retrieval sometimes returned a passage that was near the answer but not the answer, and a few bad experiences will convince a busy team that a tool is not worth the friction. We fixed the retrieval before we asked anyone to rely on it. The behavior change that mattered was teaching people to read the citation rather than the summary. Once someone confirmed for themselves that the cited page said what the answer claimed, trust followed quickly. We started with low-stakes questions, precedent lookups and 'where did we say this,' before anything touching live diligence. A couple of the analysts became the internal champions, which did more than any note from me. The lesson was that adoption is a trust problem, not a technology problem, and trust is earned one verifiable answer at a time.

What has it measurably changed?

The concrete change is that roughly 2,000 documents that were effectively unsearchable are now queryable in plain language, and a cited answer comes back in seconds rather than the better part of an afternoon. The whole deal room is searchable, not just the file someone remembered to open. I want to be careful with the framing: this is not a headcount story, and I would distrust anyone who sold it as one. What it does is move skilled time from locating information to judging it. Onboarding is faster, because a new analyst can interrogate our history directly instead of booking time with a partner for every question. The gains are real but uneven. On a document-heavy diligence sprint they are large, and on a quiet week they are modest. We measure it loosely, in hours returned and in questions that now get asked at all, because asking them finally costs seconds instead of an interruption to someone senior.

What would you say to a peer who is wary, or an investor nervous about AI touching deal data?

I would tell a wary peer to be precise about what the system does and does not do. It retrieves and cites; it does not decide. If you let people treat a fluent paragraph as a conclusion, you have built a liability, because these models will fill a gap confidently when the source material is thin. The discipline is to keep the human where the judgment is. To a limited partner or a nervous management team, the honest answer is that the confidentiality controls are stricter than the shared drive they are worried about, not looser: access is enforced by deal, everything is logged, and the data does not leave our environment or train anyone's model. The real risk is not a dramatic breach. It is quiet over-reliance, someone pasting an answer into a memo without opening the citation. We manage that with the rule that the citation must be read, and with the plain fact that the person, not the tool, signs the work.

Give me a concrete moment where it mattered.

An illustrative one, with the details changed. We were evaluating a bolt-on for a portfolio company, and a question came up about how we had structured an earn-out on a comparable services deal a couple of years earlier, and why we had set the thresholds where we did. Ordinarily that means finding whoever led that deal, or reopening a folder and reading until you locate the memo. Instead the analyst asked the question directly, and the system returned the relevant passage with a citation to the exact internal memo and page, including the reasoning we had recorded at the time. She opened the source, confirmed it, and we had our precedent in under a minute during a live call, rather than promising to circle back the next day. Nothing about that answer was invented. It was our own prior thinking, finally retrievable. That is the mundane version of the value, and it happens several times a week.

What does this mean for a firm your size, looking ahead?

For a firm our size, the strategic point is that institutional memory now compounds instead of evaporating. Every deal we do, win or lose, adds to a body of knowledge the whole team can use, which narrows the gap with much larger firms that have deeper benches and more staff to remember things. I am deliberately cautious about what comes next. There is real appetite for systems that take actions rather than just answer questions, and we will adopt more of that, but only where the output stays reviewable and reversible. Grounding and citations are not negotiable for us. The deeper change is cultural: when checking our own history costs seconds, people ask better questions and revisit their assumptions more often, which is exactly the behavior you want in diligence. The relationship with our investors and our management teams still rests on judgment and trust. AI has made our own record more useful to us. It has not, and should not, take over the parts of this business that require a person to decide.