A category thesis is not yet a company question
Most data theses among funds are category claims: public text is thinning, agents need examples of real work, established companies sit on records nobody has licensed. Each may be true, and none says which of several hundred founders, LPs and board contacts to call this week, because a category claim does not say where the flow stops.
A bottleneck does. It is the one step between "a company holds records" and "a licensed dataset is accepted" that limits everything after it. Naming it turns the category claim into a question for a specific company:
| If your thesis says the scarce thing is | The company question becomes |
|---|---|
| A CEO who wants the conversation | Who in our network can ask this CEO, and will they? |
| Records the company itself controls | Are these records the company's own, or held for its clients? |
| Records that capture how each case ended | Does the system keep each case's final state, with a date, for several years? |
| Companies large and old enough | Did this company reach 50+ full-time employees at peak, and how far back do its own records go? |
A thesis that cannot fill the left column is still a market view: useful in a fund memo, too vague for an introduction. It also tends to blur two deals; AI partnership vs data licensing separates a vendor partnership from licensing records, and only the second is what an introduction to SourceX is about.
The bottleneck-to-partner action map
Seven steps can limit a data partnership. The map sorts them by whether a network can do anything about them, then gives the partner action, the evidence that the step really binds and the evidence that would say it does not.
| Bottleneck | Can a network relieve it? | Partner action | Evidence that it binds | Evidence that it does not |
|---|---|---|---|---|
| 1. Public text running short | No; it is a market condition | Keep it as background and out of any note to a CEO | Epoch AI projects public human text could be fully used for training between 2026 and 2032 if trends hold | Epoch's own caveat that the projection assumes current trends and no major data-efficiency gains |
| 2. Records that show final outcomes | Partly; you can find companies whose systems store outcomes | Ask whether each case's end state (closed, paid, shipped, reversed) is recorded with a date | Agent evaluations score the final state of an environment, not what the agent says | SourceX's review places the record type outside current programs |
| 3. Rights to license | Partly; you can screen out client-owned material early | Ask whose records they are and leave contracts to company counsel and SourceX's rights review | Material that belongs to clients is usually not a fit | Nearly every company you ask clearly owns its operating records |
| 4. A CEO who wants the conversation | Yes; this is what a network holds | Name the relationship owner and let them ask first | Usually not a fit when the decision maker never knew about the ask | CEOs you reach say yes, then fail on fit or rights |
| 5. Size and operating history | Yes, for discovery; your network knows who is large and old | Screen for 50+ full-time employees at peak and years of the company's own records before promoting anything | Few companies in your network pass both | Most of your network passes, so the constraint sits elsewhere |
| 6. Being new to SourceX | No; history cannot be changed | Ask the CEO whether the company ever applied, held an account or exchanged email with SourceX | Earlier contact makes a company a duplicate that is not credited | Duplicates are rare once you ask before submitting |
| 7. A current program for the record type | No; only SourceX's review decides | Describe the record type and promise nothing | Relevance to current data needs is checked in review | A company outside one program fits another later |
Method note. Rows draw on two SourceX pages read on 2026-10-26, the who-qualifies page and Version 2026-10-22.2, since replaced by version 2026-11-04 of the program terms, plus two primary AI sources: Epoch AI's projection of the stock of human-generated public text and Anthropic's engineering guide to agent evaluations. "Can a network relieve it" is our judgment of what a relationship holder can change without ever handling records. The map does not measure demand for any record type, record quality, conversion or payout, and SourceX has not published those figures.
Which rows a network can move, and which it cannot
Read the map from the "Yes" rows outward.
Row 4 is the one bottleneck a venture network genuinely owns. A platform team creates neither records nor rights, but it often holds the only warm path to a CEO who would otherwise never hear about direct licensing. Name row 4 and the pilot becomes a relationship exercise: who owns each CEO relationship, and will they spend it. The platform team overview covers the program from that seat.
Row 5 is discovery: your network already knows which companies are large and old. The size test counts full-time employees at peak with contractors excluded, so a company that has since shrunk can still pass. Location sits in the same screen: Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English.
Rows 2 and 3 are partial. A partner can ask descriptive questions and rule out obvious misses, such as a company whose valuable records sit in its customers' accounts, but cannot settle either row; SourceX's review and the company's own counsel do. Nobody exports anything for that first discussion, as discussing a partnership before preparing a dataset explains.
Row 2 is the most interesting thesis and the easiest to overstate. Anthropic's guide to agent evaluations separates the outcome, the final state of the environment, from what an agent claims: a booking agent succeeds only if the reservation exists in the database. Our hypothesis, not a demand statement, is that records keeping each case's end state with a date are the kind of ground truth such tests rely on. Review, not the hypothesis, decides whether a program needs them.
Rows 1, 6 and 7 sit outside the network. Row 1 belongs in the fund memo; a CEO told that "AI is running out of data" hears a market pitch, not a reason their company fits. Row 6 is a program rule: section 3 of the terms treats any company with an earlier application, relationship, lead, account or email correspondence with SourceX as a duplicate, and SourceX checks its records before inviting an introduced company. Row 7 is SourceX's call to make, and a company turned down for one program can still match a later one.
What relationship and ownership evidence each named bottleneck needs
Each bottleneck needs different evidence before an ask, and none of it is a record. Discovery stays at the level of what exists, how much, how long, which system and who owns it.
- Access (row 4): the relationship owner's name, whether they agreed to ask, and the CEO's answer with a date.
- Size and history (row 5): peak full-time headcount with contractors excluded, the year the company's own records begin, and the operating entity that holds them today.
- Outcomes (row 2): the system of record, whether it keeps each case's end state and date, and roughly how many years of cases it covers.
- Rights (row 3): whether the records describe the company's own work or work held for clients, and whether the CEO knows of contracts that limit reuse. Counsel and SourceX's rights review decide the rest.
- Novelty (row 6): the CEO's own answer about any earlier contact with SourceX.
Who authorizes does not change by row. The CEO decides whether the company discusses licensing and signs for it; a board seat or a lead investor does not. For how the ask itself is made, see how to introduce a company for an AI data partnership.
One thesis, three bottlenecks (Illustrative)
Illustrative. A generic fund and invented details; no outcomes implied.
A fund's partners agree that mid-sized logistics operators hold years of dispatch and claims records AI developers rarely see. Three platform strategists turn that into three different pilots by naming three different bottlenecks.
| Strategist names | First five asks go to | The question asked | Result that would move the thesis |
|---|---|---|---|
| Access (row 4) | LPs and board members who know logistics CEOs personally | Would you ask the CEO whether they want a short conversation? | Owners and CEOs agree, then most companies fail on size or rights |
| Outcomes (row 2) | Portfolio founders who sell software to carriers and brokers | Among the carriers you work with, is it normal to keep closed claims with outcome dates? | Most operators keep only open cases, or purge closed ones within a year |
| Size and history (row 5) | The fund's own deal database | Which logistics companies we met reached 50+ full-time employees at peak and are more than eight years old? | Plenty pass, but nobody in the network can reach their CEOs |
Each pilot is cheap and tests one claim. Running all three at once teaches less, because a weak result cannot be traced to a step.
A narrow pilot that can show you named the wrong bottleneck
Pick one "Yes" or "Partly" row. Write three predictions before the first call, then make no more than eight asks over six weeks.
- On your row. For access: at least five of eight relationship owners agree to ask, and at least three CEOs want a conversation.
- On the next row down. Of the CEOs who agree, at least two pass every published fit fact. This catches the case where you relieved one bottleneck and ran into the next.
- On novelty. No more than one company turns out to have had earlier contact with SourceX.
Count every ask, including owners who declined and CEOs who never replied. If prediction 1 holds and prediction 2 fails, access was not your bottleneck; size and history were, so rename the thesis and run the next eight asks from row 5. If prediction 3 fails, your best-known companies may already know SourceX, and your introductions add most where the network is less visible. The thesis-to-introduction canvas helps draft the company question, and the disconfirming-evidence checklist helps judge whether a finding sinks one company or the thesis.
Before the pilot sends anyone, the person submitting signs the SourceX Channel Partner Agreement; a pilot member who signed it while it was titled Channel Partner Non-Circumvention and Non-Disclosure Agreement stays bound by that signed text. A thesis that maps the same CEO to two relationship owners still yields one credit: it belongs to the earliest valid introduction of that new company inside the 90-day attribution window and is decided when the company files a verified application.
Risks that could make the whole thesis wrong
- Data supply may not bind. Epoch AI's projection assumes current trends with no major gains in data efficiency, and names synthetic data, transfer from data-rich domains and data efficiency as ways around the limit. If those succeed, demand for licensed business records could narrow.
- Evaluations may not need your record type. Which business records a program needs is settled in SourceX's review, not by a thesis.
- Rights can fail late. A company can own its records and still face customer contracts or privacy limits that only counsel can read.
- Eight asks are a direction, not a rate. No conversion rate should be quoted from a pilot this small.
- Goodwill is spent. No data on how LPs or founders react to these asks has been published, so record it yourself.
Next step
Refer a company. No cap on rewards. Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.
When your pilot shows which bottleneck your network can actually relieve, open a partner link on the registration page, test each shortlisted company against who qualifies, and make sure every relationship owner knows the program terms before making the first ask. More thesis-to-network guides sit in the guides library.
Choosing a bottleneck for a referral pilot is a network decision, and judging SourceX itself as an investment is a different one; readers weighing the second can request investor materials through the SourceX investor overview.
This article is for information only and is not an offer to sell or a solicitation of an offer to buy securities. Any offering by SourceX is made only to verified accredited investors through its offering documents.