# Can better introductions improve AI data? | SourceX

Canonical: https://refer.sourcex.si/guides/how-better-business-introductions-could-improve-the-data-behind-ai
Last updated: 2026-11-04

Mostly through context, not volume. A trusted introduction can reach the CEO who can authorize a license, name the system and the person who runs it, and say which records the company itself owns. That context may make a dataset review faster and its provenance clearer. It cannot, on its own, show that a model trained on the data performs better: that is a separate technical test on the delivered dataset. The seven-link pathway below keeps the two claims apart and names where each link can break.

The temptation is to draw a straight line from "I know this company well" to "AI gets better data". The line exists, but it has seven joints, and a partner can see only the first three.

## The information a context-free lead loses

A purchased list or a cold contact gives a reviewer a row: sector, approximate headcount, a name with a job title. Almost everything a dataset review needs first is missing.

| Context | Context-free lead | Trusted introduction | Why a review needs it |
|---|---|---|---|
| Who can say yes | A job title, often not the CEO | The CEO, who knows the request is coming | If the person who decides is unaware of it, the introduction usually counts as no fit |
| Where the records live | Unknown | Named systems and the person who administers them | Decides how records could be reached and prepared |
| How far back they go | Unknown | A span of years per system | Operating history is part of fit |
| Whose material it is | Unknown | Company-owned versus client-owned, at least as the CEO understands it | Client-owned material is usually not a fit |
| Why the records exist | Unknown | The workflow that produced them | The first question a dataset's documentation has to answer |
| What is changing | Unknown | A dated migration or archive change, if any | Real timing, not invented urgency |

The cost of lost context in AI data is documented outside this program. The Data Provenance Initiative audited 1,858 text datasets and reported license omission rates above 70% and license error rates above 50% on widely used dataset hosting sites. Those datasets lost their context somewhere between the people who made them and the people who used them. A licensing path that starts with the company's own CEO starts with that context intact.

## How a trusted introduction surfaces the owner and the workflow

Two things travel with a good introduction that no database holds. The first is authorization: the CEO hears about licensing from someone they already trust, decides whether to discuss it and, under the program, authorizes the discussion and signs. The second is workflow knowledge: an accountant knows which system closes the books, an IT provider knows which helpdesk holds ten years of tickets, an operating partner knows the company migrated its CRM in 2021.

Both reach SourceX only at metadata level. Partners do not need to export, upload or describe confidential records, and an IT provider that administers a company's systems must not use that access to look at or sample records for an introduction. What it shares is what the CEO approves: the system names, record families, years and administrators. The [metadata-only opportunity map](/guides/metadata-only-referral-opportunity-map) gives the columns, and the [managed service provider page](/for/managed-service-providers) covers that partner type.

## Why clearer provenance may make review more efficient

This is the hypothesis, stated as one. Datasheets for Datasets, a proposal published in Communications of the ACM, argues that every dataset should come with documentation of its motivation, composition, collection process and recommended uses, to improve transparency and help practitioners choose the right dataset. An introduction cannot write a datasheet, but it can seed the first answers at metadata level, which may spare a review several rounds of basic questions.

| Datasheet question area | What an introduction can supply | What only review can establish |
|---|---|---|
| Motivation: why the records exist | The workflow that produces them | Whether that workflow matches a current data need |
| Composition: what the records are | Record families, systems, rough counts | Actual content, quality and duplication |
| Collection process: how they were made | Who enters them, in which system, since when | Gaps, format changes and consistency over time |
| Uses and restrictions | Whose material it is, as the CEO understands it | Rights under customer contracts, privacy and third-party terms |
| Distribution and maintenance | Who administers the system | Scope, preparation and handoff under a signed agreement |

Whether this actually shortens reviews is not yet measured. SourceX has not published review-efficiency figures by introduction type.

## The separate test that decides usefulness

A dataset's usefulness for AI is measured by training or evaluating models on it, not by how it was sourced. DataComp shows what such a test looks like: it holds training code and compute fixed so that only the training set varies, then evaluates each resulting model on 38 downstream test sets. Whoever trains on licensed records runs its own version of that comparison, and the results belong to them.

So the honest claim for a partner stops before that test. "My introduction reached the CEO, named the records and surfaced a rights question early" can be true and checked. "My introduction improved a model" cannot be checked by the partner and should never be made. The [data quality page](/proof/why-data-quality-beats-quantity) explains why documented, real records attract interest; the [trajectory evaluation gap analysis](/questions/trajectory-level-evaluation-evidence-gaps) shows how missing context limits what can be graded later.

## The introduction-to-data-quality causal pathway

This is the asset. Each link is a mechanism with an observable sign and a named failure. A chain is only as strong as its weakest link, and a break anywhere stops the claim at that point.

| Link | Mechanism | Observable sign | Where it breaks | Who can observe it |
|---|---|---|---|---|
| 1 | A trusted relationship brings the request to the CEO | The CEO knew in advance and agreed | Introduction goes around the CEO | Partner |
| 2 | The CEO names the right records and their administrator | Systems, years and an administrator are named | The CEO does not know where records live, or an outside vendor holds them | Partner |
| 3 | Named records come with an ownership picture | Company-owned versus client-owned is stated | The richest records belong to clients | Partner, then SourceX |
| 4 | Clear context at intake reduces basic review questions | Fewer clarification rounds | The partner's account was wrong, and review spends time correcting it | SourceX |
| 5 | Review turns into an agreement and preparation plan | Scope and preparation agreed, personal details handled | Rights review narrows scope to little | SourceX and the company |
| 6 | The prepared dataset carries documented provenance | Motivation, collection and restrictions documented | Documentation stays thin | SourceX |
| 7 | Documented records prove useful in a model | A downstream evaluation with training held fixed | No measurable gain | Whoever trains on the data |

The graphic shows the same seven links in three lanes, with a wall before link 7.

## One relationship followed through, and one that stops (Illustrative)

Illustrative. Generic companies, invented facts, no outcomes implied.

**A 340-person property management company**, introduced by its managed IT provider. Link 1 holds: the provider's account lead has known the CEO for years and asks first. Link 2 holds: work orders in one system since 2012, tenant messages in a helpdesk, both administered by the provider. Link 3 is mixed: work orders are the company's own, while tenant messages contain personal details that would have to be removed. The CEO approves sharing that picture, and the provider submits the introduction. Links 4 to 6 are now SourceX's and the company's to test; the outcome is open. Link 7 will never be the provider's claim, whatever happens.

**A 90-person call center outsourcer**, introduced by its accountant. Links 1 and 2 hold: the CEO agrees, and recordings and QA scorecards go back eight years. Link 3 breaks: the recordings belong to the outsourcer's clients under its service contracts, and the who-qualifies page treats client material as usually not a fit. The chain stops. The introduction still did its job by reaching that answer in one conversation, with nothing exported and no records seen.

## What partners and SourceX could measure without overstating AI impact

| Measure | Link | Definition | Held by | Not evidence of |
|---|---|---|---|---|
| CEO-consented introductions | 1 | Share of introductions where the CEO knew and agreed beforehand | Partner | Fit |
| Facts known at handoff | 2 | Of size, location, own records and authorized contact, how many were known | Partner | Record quality |
| Early rights flags | 3 | Introductions where an ownership question was raised before submission | Partner | A rights decision |
| Clarification rounds | 4 | Basic questions a review had to ask before assessing the records | SourceX; not published | Dataset usefulness |
| Scope after rights review | 5 | How much of the named records stayed in scope | SourceX; not published | Model performance |
| Provenance completeness | 6 | Share of documentation fields answered for a prepared dataset | SourceX; not published | Model performance |
| Downstream usefulness | 7 | Evaluation results with training held fixed | The developer; not published | Anything about the introduction alone |

None of these figures has been published. A partner can start on the first three tomorrow; the middle three would need SourceX to publish them with cohorts and definitions; the last one is not SourceX's or the partner's to report.

## What every introduction still has to meet

Context helps a review see fit; it does not create it. 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. Beyond location, the company must have reached 50+ full-time employees at its peak, contractors aside, must document its operating history and must hold the rights to license its own records, and the CEO authorizes the discussion and signs for the company. Material belonging to clients, personal consumer data and one-off files with no documentation or context are usually not a fit. Two things stay open until SourceX reviews the company: whether its records fit current demand, and what a license would cover. Partners sign the SourceX Channel Partner Agreement, formerly titled the Channel Partner Non-Circumvention and Non-Disclosure Agreement, before submitting introductions. Payment follows the conditions in the terms and is due within 60 calendar days after the later of two dates: the day the Acquisition Partner's final written acceptance of the complete, validated dataset is given, and the day handoff is completed.

## 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.

If you can hold links 1 to 3 for a company you know, [sign up as a referral partner](/register), compare that company with [SourceX's who-qualifies criteria](/who-qualifies), review [the program terms](/terms), and then pass on the context the CEO approves and nothing more.

## Sources

- https://refer.sourcex.si/terms
- https://refer.sourcex.si/who-qualifies
- https://refer.sourcex.si/how-it-works
- https://cacm.acm.org/research/datasheets-for-datasets/
- https://arxiv.org/abs/2310.16787
- https://arxiv.org/abs/2304.14108
