- A decade of deal folders is an asset only if it can be queried, and most firms hold one that nobody has opened in eighteen months.
- A shared drive is not a dataset: much of it is scanned rather than text, numbers and their meaning sit in unlinked files, entities are unresolved across naming conventions, and nothing records what actually happened.
- Ingestion has five stages — inventory, extraction, entity resolution, outcome labelling and indexing — and the middle three are where the time goes.
- Between 20% and 40% of a long-lived deal drive is duplicates and superseded versions, and loading all of it degrades every subsequent answer.
- Outcome labelling — recording whether each deal closed, passed, or underperformed — is manual, fast, and the highest-return hour a firm spends on the project.
- Before ingesting, establish in writing where the data lives, that nothing is used to train a model, what structured form comes back out on termination, and what a re-run costs.
Every firm we speak to has this folder. Sometimes it is Dropbox, sometimes SharePoint, sometimes a Box instance nobody has audited since the last IT contractor left. Inside it is every data room the firm has ever downloaded, every memo it has written, every model, every closing binder, and a substantial quantity of files named Copy of Copy of NDA_v3_FINAL.pdf.
Two beliefs are commonly held about this folder at the same time. The first is that it is the firm's most valuable proprietary asset. The second is that nobody has opened anything in it in eighteen months. Both are true, and the gap between them is the whole subject.
Why it is not a dataset
The instinct is to point an AI tool at the folder and start asking questions. This works badly, and it is worth understanding why, because the failure is not obvious in a demo.
It is not text. A meaningful share of what is in there is scanned. Contracts signed in 2016, third-party reports delivered as image PDFs, hand-annotated schedules photographed and filed. Anything that reads the folder without optical character recognition will silently skip these, and the skipping is silent — you get answers, they just do not include the documents that were pictures.
The numbers are in spreadsheets and the meaning is in prose, and nothing connects them. The model has the customer concentration figure. The memo has the argument about why it was acceptable. Nothing in the file system links the two, and a retrieval system that treats each file as an independent document will return one or the other and not know it has half an answer.
Entities are not resolved. The same company appears as Meridian Industrial, Meridian Industrial Holdings, MIH, and Project Falcon, depending on which stage of the process the document came from. Until those are the same entity in the system, no question about that company returns a complete answer, and — worse — none of the answers announce their incompleteness.
There is no ground truth on outcome. The folder contains what the firm thought. It rarely contains what happened. A memo predicting 22% EBITDA growth is not labelled with the fact that the business delivered 11%. Without that, the archive can tell you what you believed and cannot tell you when you were wrong, which is most of the value.
What ingestion actually involves
The honest version of this project has five stages, and the middle three are where the time goes.
Inventory. Before anything is loaded, someone has to establish what is in there, in what formats, spanning what period, and which of it is worth carrying forward. In our experience 20% to 40% of a long-lived deal drive is duplicates and superseded versions, and loading all of it degrades every subsequent answer by making the system confident about a stale document.
Extraction. Text out of scans, tables out of PDFs, structure out of spreadsheets where the structure is implied by formatting rather than declared. This is the part that determines the quality ceiling of everything downstream, and it is the part most vendor demos skip by using clean sample data.
Entity resolution. Deciding that Project Falcon and Meridian Industrial Holdings are the same company, that the CFO named in the 2019 management presentation is the same person as the one named in a 2023 CIM for a different target, and that two documents describing the same supplier are describing the same supplier.
Outcome labelling. Attaching to each deal what actually happened: closed, passed, lost at auction, closed and underperformed, closed and exited at 3.1x. This is manual, it is fast, and it is the single highest-return hour a firm spends on the project.
Indexing for the questions you will actually ask. Not general search. The specific queries a deal team runs: have we seen this business, has this management team appeared before, what did we conclude about this sector in 2022, which of our passes cited a reason that also applies here.
How long, honestly
For a mid-market firm with ten to fifteen years of history and a few hundred deals, expect the technical work to run in weeks rather than months, and expect the inventory and outcome-labelling decisions to be the schedule driver, because they need a partner's attention and partners are busy.
Anyone who tells you a decade of unstructured deal folders is a drag-and-drop is describing the upload step and calling it the project.
The questions to ask before you start
Where does the data live once it is ingested, and who can see it? Historical deal files contain terms, LP information, and material non-public information about companies that are now owned by somebody else. This is the most sensitive corpus the firm holds, and it is about to be moved.
Is anything in this used to train a model? The answer must be no, in the contract, not in the FAQ.
What comes back out if we leave? If the extracted, structured, entity-resolved version of your archive is only accessible inside a vendor's interface, the vendor now holds the improved version of your own asset. Ask for the export format, in writing, before ingestion rather than after.
What does it cost to re-run? Extraction quality improves over time. A firm that has to pay full ingestion cost again in two years to benefit from that will not do it.
What you get on the other side
The change is not that search gets better. It is that a question that previously required a partner to be in the room becomes answerable by anyone on the deal team in the first week.
Have we seen this company before? — including under a different project name, including as a competitor named in somebody else's CIM.
Have we seen this management team before? — including in a business that did not go well.
What did we conclude about this sector the last four times? — including the passes, which is where the reasoning is densest.
Has this failure mode appeared in our history? — the question that only becomes askable once the archive knows what a failure mode is.
None of that is exotic. It is the firm's own knowledge, returned to the firm at the moment it is needed rather than eighteen months later in a conversation that begins with somebody saying "didn't we look at one of these?"
Bulk ingestion of historical deal folders is a standard part of a HuxleyIQ deployment; we will walk through the inventory step against your own drive before anything moves.