# Venture platform AI contribution: who does what | SourceX

Canonical: https://refer.sourcex.si/guides/a-venture-platform-can-help-build-better-ai-without-training-models
Last updated: 2026-11-04

A venture platform helps build better AI by reaching companies whose ordinary records show real work and how it turned out, and by asking their CEOs whether they want to license those records. The platform supplies relationships and metadata. SourceX handles rights review, preparation and the license. AI developers turn licensed records into training sets, evaluation tasks and graders. Place each contribution in one of those three roles before you describe any of it to founders or LPs.

That is a smaller claim than "our portfolio is powering AI", and a more useful one: it tells a platform team which companies to ask about, which questions to ask and which sentences never to write.

## What AI developers need that only operating companies hold

Two primary sources describe what evaluating AI on real work requires. Anthropic's engineering guide to agent evaluations defines a task as a single test with defined inputs and success criteria, and the outcome as the final state of the environment, which can differ from what the agent reports; its example is a flight-booking agent whose success depends on whether a reservation actually exists in the database. OpenAI researchers' GDPval benchmark builds its tasks from the work of industry professionals with about 14 years of experience on average, across 44 occupations in the nine largest U.S. GDP sectors, and grades mainly by blinded comparison with expert work because grading such tasks automatically is hard.

Read from a platform's seat, the scarce ingredients are realistic inputs, a recorded end state and an experienced person's judgment. Operating companies produce all three as a byproduct of doing business: a request arrives, someone follows a procedure, something happens, and often someone checks the result.

| What an evaluation needs | Where the source describes it | Where it sits in an operating company | Record families (Illustrative) |
|---|---|---|---|
| Realistic inputs | A task's defined inputs (Anthropic) | Requests the company receives | Service tickets, work orders, claim submissions |
| Success criteria | A task's success criteria (Anthropic) | Procedures and specifications | SOPs, QA rubrics, contract specifications |
| A recorded end state | The outcome as final state (Anthropic) | What actually happened | Resolution codes, settled amounts, repeat-visit flags |
| Expert judgment | Comparison with experienced professionals (GDPval) | Review by senior staff | Reviewer edits, approvals, QA scores |

A platform's leverage is knowing which companies keep those four together, in one system, for years. For the plain-language version of the training and evaluation distinction, see [training versus evaluation, briefly](/guides/how-to-explain-the-difference-between-training-and-evaluation-briefly).

## The introduction contribution map

Six contributions sit between a company's records and a better model. The platform makes two of them, shares a third with the CEO and makes none of the rest.

| Contribution | Who makes it | What the platform does | What the platform may say | What it never says |
|---|---|---|---|---|
| 1. Find companies that keep requests, procedures, outcomes and reviews together | Platform | Screens its network from what relationship owners already know | "We identified six companies with linked request and outcome records" | That those records are valuable |
| 2. Ask the CEO with permission | Platform, through the relationship owner | Asks once and records yes, not now or no | "Three CEOs chose a licensing conversation" | That a CEO was persuaded |
| 3. Describe the records | Platform, with the CEO's approval | Passes system, record families, years, administrator and ownership as the CEO understands it | "Each introduction carried a metadata summary" | Anything about content, which it never sees |
| 4. Review rights, agree terms, prepare and deliver | SourceX and the company | Steps back | Nothing | Any part of the license |
| 5. Build tasks, success checks and graders; train models | AI developer | Nothing | Nothing | "Our portfolio trains models" |
| 6. Measure whether a model improved | AI developer | Nothing | Nothing | "Our introductions improved a model" |

The counts in column four are Illustrative. The sentences in column five are the ones that creep into LP letters and founder newsletters; none of them can be checked by the person writing them, and the last one names a result only a developer can observe. SourceX does not name its Acquisition Partners, so a platform cannot say which developer used anything either.

**Method note.** The map divides the work using SourceX's description of its own role on its investor page (sourcing, rights review, contracting, preparation, delivery and payment) and the two evaluation sources above, checked 2026-10-26. It counts companies, consents and metadata summaries. It does not measure the value of any dataset, model performance or payout. How the context in an introduction may make a review clearer is a separate question, worked through link by link in [how better introductions could improve the data behind AI](/guides/how-better-business-introductions-could-improve-the-data-behind-ai).

## Translate the contribution thesis into a company question

A platform needs a thesis it can lose. A working example, not a finding:

> **Thesis (proposed).** Our network contributes most through companies that link a request to its outcome and to a reviewer's judgment, not through the companies with the most records.

Volume is the easiest trait to claim and the hardest for an outsider to observe. Linkage is a yes-or-no question a CEO can answer in a minute, so the thesis becomes:

**Company question.** Does this company keep, in one system it controls, what was asked, what was done, how it turned out and who checked it, going back five or more years?

The question deliberately leaves out whether the company uses AI, how fast it is growing and how it is funded. None of those appear in the published fit criteria, and none tells you whether the four ingredients exist.

## The ownership and relationship facts every contribution rests on

Contribution 2 needs a person and contribution 3 needs a clean ownership picture, and the map is only as good as those two.

- **The relationship owner.** Someone in your network who can ask the CEO and be heard, and who has agreed to. Owning shares does not make the fund that person.
- **The CEO.** Opening a licensing discussion is the CEO's decision, and so is the signature. Operations leaders can describe systems; they do not authorize.
- **Outcome records.** These are where ownership gets complicated. Settlement amounts may sit in files the company holds for insurers; repeat-visit flags carry customer names. Material that belongs to a company's clients is usually not a fit, and personal details are removed in SourceX's preparation workflow, but the platform settles neither question; SourceX's rights review does.

The floor for any introduction: 50 or more people on the full-time payroll at peak, contractors aside, plus documented operating history and rights to license. 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. Before asking, confirm the company has no earlier lead, account or email thread with SourceX; the terms treat any such contact as an existing relationship, so the company would be reviewed as a duplicate. Submissions also wait on a signed SourceX Channel Partner Agreement, the current name of the document once called the Channel Partner Non-Circumvention and Non-Disclosure Agreement. Fit to a current program, any reuse limits in customer contracts and the scope of a license are decided only in SourceX's review.

## A small test: eight relationship owners, one company each

Test the thesis before it shapes a platform plan.

1. **Pick eight companies**, one per relationship owner, from across the network, not only the portfolio.
2. **Predict first.** Each owner marks the four ingredients for their company from memory: present, absent or unknown.
3. **Ask the CEO.** With permission, the owner asks the same four questions at metadata level, then whether the CEO wants a licensing conversation.
4. **Record three counts:** predictions the CEO confirmed, companies with all four ingredients linked, and CEOs who chose an introduction.

**What would change the thesis, decided in advance.**

| Result | Reading | What changes |
|---|---|---|
| Fewer than three of eight keep request and outcome linked | Linked records are rare in this network | Drop the thesis; your contribution, if any, is reach, not record type |
| Owners mispredict ownership or linkage for more than two of eight | Memory is not evidence | Describe no contribution until a CEO conversation has confirmed it |
| Linkage is common but CEOs decline | The contribution exists on paper | Stop asking and revisit on dated events such as a system migration |
| Linkage common and CEOs say yes | The thesis survives eight companies, which proves little | Continue with consented introductions and keep counting |

SourceX has not published conversion rates or results from tests like this one, so the readings compare your network only with the thresholds you wrote. A first-month plan for running the asks themselves is in the [first-month referral plan for venture platform teams](/guides/a-first-month-referral-plan-for-venture-platform-teams).

## Where the contribution story overreaches

- **Developers decide what they need.** Tasks, graders and harnesses are their design. A record family the platform prizes may not suit any current program; SourceX checks relevance in review, and criteria differ by program.
- **Automatic grading could spread.** GDPval relies on expert comparison because these tasks are hard to grade automatically. If that changes for a domain, expert review records there matter less.
- **Rights narrow the best records.** Outcome data is the ingredient most likely to contain third-party details or belong to clients.
- **Founder attention is finite.** Every ask spends a relationship owner's goodwill, whatever the thesis says.
- **Reporting drift.** "Three CEOs chose a conversation" is a fact; "we are helping build frontier AI" is a slogan. Keep platform reports to the first kind. How platform teams explain the program to founders is covered in [explaining company data partnerships](/guides/how-venture-platform-teams-can-explain-company-data-partnerships).

Licenses can cover evaluation only; the [evaluation-only license comparison](/compare/evaluation-only-data-license) explains what that changes for the company. Program basics for funds sit on the [venture platform teams page](/for/venture-platform-teams).

## 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 company on your map keeps its requests, outcomes and reviews together and its CEO wants the conversation, [request your partner link](/register), confirm that it clears [who qualifies](/who-qualifies) and read the [program terms](/terms) in full before the relationship owner writes the introduction.

A platform's referral work and an assessment of SourceX as a possible investment are two different exercises; for the latter, investor materials can be requested 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://www.anthropic.com/engineering/demystifying-evals-for-ai-agents
- https://arxiv.org/abs/2510.04374
- https://refer.sourcex.si/terms
- https://refer.sourcex.si/who-qualifies
- https://sourcex.si/investors
