# Vertical AI fund referral pilot: one-page charter | SourceX

Canonical: https://refer.sourcex.si/guides/a-referral-pilot-for-a-fund-exploring-vertical-ai
Last updated: 2026-11-05

A fund exploring vertical AI can test its thesis with a narrow referral pilot instead of a portfolio-wide push. Pick one sector, write the prediction that its established operators keep years of their own decision records, name the result that would disprove it, and ask relationship owners about no more than twelve companies over eight weeks. Aim at the incumbents in the sector, not the fund's AI startups, which rarely pass the size and history tests. Settle the conflict with your own portfolio before the first call.

The pilot is a referral experiment and nothing more. It says nothing about what any company's records are worth to an AI developer; only SourceX's review and a signed agreement can establish that.

## From a sector thesis to a question one company can answer

A vertical AI thesis usually runs like this: products built for one industry need that industry's records, the forms, exceptions, decisions and outcomes that general models rarely see. There is evidence on both sides. Bloomberg's researchers trained BloombergGPT, a 50-billion-parameter model, on a 363-billion-token dataset built from Bloomberg's own financial sources plus 345 billion tokens of general data, and reported that it beat existing models on financial tasks without losing ground on general benchmarks. That is the company's own result, but it shows a developer treating sector records as worth building around. NIST's AI Risk Management Framework adds a reason sector developers care about provenance: datasets used to train AI systems may become detached from their original context, or stale relative to where the system is deployed.

A thesis at that altitude cannot be tested by a platform team. Bring it down to one company:

> Does this 300-person commercial claims administrator keep a decade of adjuster decisions with recorded outcomes, in a system it controls, with the right to license them, and does its CEO want to discuss it?

One relationship owner can answer that with yes, no or unknown. Twelve such answers test the thesis better than a sector report, and the [vertical AI training data guide](/guides/vertical-ai-training-data) helps choose which sector's records to ask about.

## The vertical thesis pilot charter

Write the charter on one page before anyone is contacted. The example column is Illustrative.

| Field | What the fund writes | Example (Illustrative): commercial claims |
|---|---|---|
| 1. Sector boundary | One segment, defined by the work companies do | Third-party claims administrators and self-insured employers' claims units in the U.S. and Canada |
| 2. Prediction | One sentence a company can contradict | Operators with 150+ people keep 8+ years of claim files with final outcomes in a system they control |
| 3. Disconfirming result | The count, and the week, that narrows or ends the thesis | Fewer than 3 of 12 companies own their claim files, by week 6 |
| 4. Population | Up to 12 companies, each named by a relationship owner | 12, named by four partners and two LPs |
| 5. Fit facts before any ask | The published tests, each marked known, unknown or failing | Size at peak, operating history, rights to license, location |
| 6. Path | The relationship owner who asks, and the CEO who decides | One owner per row; the CEO is the only authorizer |
| 7. Conflict rule | How portfolio relationships are protected | No portfolio company's customer is approached without that founder's written yes |
| 8. Information rule | What may be collected | Record type, rough count, years, system and owner; no files or samples |
| 9. Stop date | The date and the three possible decisions | Week 8: widen to a second sector, repeat narrowed, or stop |
| 10. Report | What the write-up contains | Counts and reasons per row; no reward forecast |

**Method note.** Fields 4 to 8 come from the program terms (version 2026-10-22.2, since replaced by version 2026-11-04), the who-qualifies page and the sample Channel Partner Agreement, read on 2026-10-26. Fields 1 to 3 and the thresholds are the fund's own design; ours are examples. The charter measures relationship paths, fit facts and ownership, not demand for any sector's records, which SourceX checks in review. SourceX has published no pilot results, mandate feedback or founder interviews, so a fund's first run produces that evidence for itself.

## Relationship and ownership evidence for each named company

Every row needs three kinds of evidence, each graded known (stated by the CEO or the relationship owner), inferred (read from public material, never acted on alone) or unknown.

**The path.** Who owns the CEO relationship: a partner, an LP, a board member or a portfolio founder. Whether the CEO knows the question is coming. Only the CEO can authorize a licensing discussion, and whoever submits must first have signed the SourceX Channel Partner Agreement (a copy the fund signed while the agreement was titled Channel Partner Non-Circumvention and Non-Disclosure Agreement stays in force as signed).

**Ownership.** Whose records they are. In outsourced sectors this is the crux: an administrator's claim files may belong to the carriers and employers it serves, and under SourceX's screening guidance, records owned by the clients a company serves are usually not a fit. Customer contracts and SourceX's rights review decide it; the fund only records what the CEO says.

**The fit facts.** A peak of 50 or more full-time employees, contractors excluded; years of records the company produced itself; rights to license; and location. 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. Funding plays no part in any of these tests.

For each row, credit follows the program's attribution rule: of all valid introductions of a new company, only the first inside the 90-day window counts; it is credited once the company submits a verified application, and a company has only one credited partner. A company that already had any lead, account or email exchange with SourceX is reviewed as a duplicate. The [map of founder-owned companies around a fund](/guides/a-practical-map-of-founder-owned-companies-around-a-vc-fund) helps find the owners for each row.

## The conflict a vertical AI fund has to settle first

A fund that backs vertical AI startups is about to approach the incumbents those startups sell to. Three tensions follow, and the charter's conflict rule has to answer all of them.

1. **The relationship belongs to the founder.** Asking a portfolio company's customer about its records spends that founder's goodwill. Get the founder's written yes, or drop the row.
2. **Where records go is not visible.** SourceX's sample partner agreement says an Acquisition Partner, which may be an AI lab, a large enterprise AI developer or an organization working with them, can keep its identity and contract confidential, and SourceX need not reveal Acquisition Partner names, contracts or prices to partners. The fund cannot know whether a licensed archive reaches a competitor of its own portfolio company, and must not suggest to anyone that it can.
3. **The fund cannot steer.** SourceX deals with the company directly and the company sets its own scope and terms. A partner never handles the records.

A simple test: if a portfolio founder would object to a customer licensing records to an unnamed AI developer, leave that customer off the list.

## An eight-week run with its stop rules written first (Illustrative)

Illustrative. A generic fund, invented counts and no outcomes implied.

- **Week 0.** The platform lead and the partner who holds the thesis sign the charter.
- **Weeks 1 and 2.** Relationship owners confirm their path and mark the fit facts they know for each of the 12 rows.
- **Weeks 3 to 6.** Owners hold metadata conversations and ask each CEO whether a conversation with SourceX would be welcome.
- **Week 6.** The disconfirming count is checked.
- **Week 8.** Decision and write-up.

| Result at week 8 (Illustrative) | Companies of 12 |
|---|---|
| Path confirmed by the relationship owner | 9 |
| All four fit facts known and passing | 5 |
| Claim files owned by the company, not its clients | 2 |
| CEO wanted a conversation | 3 |
| Introduced in writing | 2 |

Only 2 of 12 companies owned their claim files, below the threshold of 3, so the ownership prediction failed at week 6. The fund does not widen. It narrows the sector to self-insured employers' claims units, whose files belong to the employer, and repeats with a new list of up to 12. A pilot that ends with "the records sit with the carriers, not the administrators" has done its job. For a fuller interview protocol, see [testing a data supply thesis through founder conversations](/guides/how-to-test-a-data-supply-thesis-through-founder-conversations); for turning AI research into one question per founder, [what a new data category means for your founder network](/guides/what-a-new-data-category-means-for-your-founder-network).

## Findings that should change or end the thesis

- **General models may close the gap.** Nori and colleagues reported that GPT-4, steered only by general-purpose prompting they called Medprompt, set state-of-the-art results on all nine MultiMedQA benchmark datasets and beat the specialist Med-PaLM 2 by a significant margin. If prompting a general model matches sector-specific training in your vertical, developers may need fewer sector records than the thesis assumes.
- **The records belong to someone else.** In outsourced sectors, files often belong to the clients served, which the program usually treats as not a fit. The Illustrative run above failed on exactly this.
- **The operators are small.** If most companies in the sector stay below the 50-employee line at peak, the population cannot fill a charter.
- **Old context.** NIST's warning about stale data cuts both ways: an archive of workflows the sector has abandoned may matter less than a recent one.
- **Demand is decided in review.** SourceX checks each company against its current programs, and criteria differ by program. A row that passes every fit fact can still be declined.
- **One company's evidence.** BloombergGPT's results are reported by its builders; treat them as one data point, not proof that every sector's records are in demand.

**Methods note for the risk list.** Each risk is drawn from the cited research and SourceX's published criteria as read on 2026-10-26. None was measured in a pilot, and none says how many companies in any sector would pass.

## 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 a charter row shows a confirmed owner, four known fit facts and a CEO who wants the conversation, [create your partner link](/register), set the row beside [who qualifies](/who-qualifies) one last time and work through the [program terms](/terms); the portfolio-wide view is on [the page for venture platform teams](/for/venture-platform-teams), and more in the [guides library](/guides).

Running a sector pilot is a referral decision; diligence on SourceX as a company is a different one, and a fund that wants to start it can request investor materials, naming the diligence questions it needs answered, through the [SourceX investor overview](https://sourcex.si/investors).

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.

## Sources

- https://refer.sourcex.si/terms
- https://refer.sourcex.si/who-qualifies
- https://sourcex.si/documents/partner/SourceX_Channel_Partner_Agreement_Sample.pdf
- https://arxiv.org/abs/2303.17564
- https://arxiv.org/abs/2311.16452
- https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf
- https://sourcex.si/investors
