Skip to content

Rights and contracts

Enterprise data rights gap survey: how to measure what stops deals

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

This is a protocol, not a result. A study can measure the rights gaps that stop data licensing deals by sampling candidate datasets inside companies that considered licensing, then coding the actual contracts, notices and terms behind each one against a counsel-reviewed codebook of ten gap types. Each gap is rated blocking, fixable or not material. No findings are reported here.

License terms

Enterprise data rights gap

This page covers enterprise data rights gap survey: how to measure what stops deals.Illustrative

Key takeaways

  • This page publishes a study method and reports no results.
  • Sample company-dataset pairs, not companies, because rights positions differ between record families in one business.
  • Code the governing documents, including the notice in force when records were created, rather than relying on what respondents remember.
  • Rate each of ten gap types blocking, fixable, not material or undetermined under counsel-reviewed rules.
  • Compare datasets that reached signature with those that stopped, and compare stated stop reasons with coded evidence.

How could a study measure the data rights gaps that stop licensing deals?

Measure gaps in documents, not in opinions. The study described here samples candidate datasets, meaning a defined set of records a company considered licensing, and for each one collects two things: a short structured interview with the person who owns the decision, and the documents that govern the records, such as customer contracts, privacy notices, software terms, contractor agreements and prior licenses. Trained coders then apply a fixed codebook that names ten kinds of rights gap and rates each as blocking, fixable or not material.

The headline measure is simple to state: of the candidate datasets assessed, what share had at least one blocking gap, and which gap types appear most often among them. Every estimate carries its denominator, its sampling method and an interval. This page publishes the design, the instruments and the codebook so the method can be reviewed and preregistered before anyone collects data. It reports no results, and none of the categories or examples below should be read as a finding.

Which problems count as a rights gap in this study?

A rights gap is a missing permission for the specific use under study: licensing the records, usually after de-identification, to a third party that will train or evaluate AI models. The definition deliberately keeps four questions apart, because companies often answer one and assume they have answered all four.

Access is whether the company can export the records at all. Privacy permission is whether the notices and law that applied when personal data was collected allow this new use. Ownership is whether the company, rather than a vendor, contractor or customer, holds the underlying rights. Contractual permission is whether customer contracts, platform terms and earlier deals allow licensing. A company can have full access and still fail on the other three, and de-identification on its own does not supply any of them.

Problems that are not rights gaps are coded separately so they cannot inflate the count: poor data quality, too little history, no buyer interest, internal reluctance and price disagreements. The study records them as competing reasons a deal stopped, which lets the analysis separate deals that died on rights from deals that died for other reasons.

Who and what gets sampled?

The unit of analysis is the company-dataset pair, not the company. A company may consider licensing its support tickets and its engineering tickets, and the two can have entirely different rights positions. Each sampled company nominates up to three candidate datasets from a fixed list of record families, and if it names more, the study selects three at random so respondents cannot steer toward the clean ones.

The target population is operating companies above a stated employee floor, in named countries and sectors, that have considered licensing records within a stated look-back period. A list frame is bought or built and its coverage problems are disclosed. For U.S. strata, the Census Bureau's Statistics of U.S. Businesses gives counts of firms by industry and enterprise employment size, which the study uses to set stratum sizes and to weight results back to the population, not as a contact list. AAPOR's standard definitions for establishment surveys guide how cases are classified when a business cannot be reached, has closed or is out of scope.

Screening is the costly step, because only some companies have considered licensing at all. The screener therefore asks one factual question first, whether anyone at the company evaluated sharing or licensing a set of records with an outside party for AI use in the period, and records every screened-out case so the eligibility rate itself becomes a reported quantity.

What do the interview and the document request collect?

The interview captures context the documents cannot: what was proposed, how far it got and why it stopped. The document request captures the evidence the codebook needs. Both are fixed before fieldwork and versioned.

What do the interview and the document request collect?
BlockCollected throughExample itemsWhy it is needed
Dataset profileInterviewRecord family, source systems, years covered, whether personal data is presentDefines the denominator and strata
Deal stageInterviewConsidered, assessed, offered, negotiated, signed, deliveredSeparates early drop-offs from late failures
Stated stop reasonInterview, open text then codedRespondent's own account of why the deal stoppedCompared later with the document coding
Governing documentsDocument request under confidentiality termsCustomer contract templates, privacy notices in force at collection, SaaS terms, contractor agreements, earlier licensesThe evidence coders actually code
Remediation attemptsInterviewConsents sought, contracts amended, records excludedDistinguishes fixable from blocking gaps
FirmographicsFrame plus interviewSector, employee band, country, ownership typeWeighting and subgroup analysis

The evidence codebook: ten gap types and how each is rated

The codebook is the core asset. Each code names a gap, the evidence that triggers it and the rule for rating it. Counsel reviews the definitions and rating rules before fieldwork; coders apply them without making legal judgments about any individual company.

At a glance

The evidence codebook: ten gap types and how each is rated

Two coders code every dataset independently, a third adjudicates disagreements, and counsel audits a random sample of adjudicated codings against the rating rules. Agreement is reported per code using a chance-corrected statistic chosen in the preregistration, with a minimum threshold set in advance. A code that misses the threshold in the pilot is rewritten and re-piloted, not dropped quietly.

Counsel's role is bounded. Counsel reviews code definitions and rating rules for legal accuracy, writes the decision rules that G04 and G08 depend on, and audits the sample. Counsel does not issue an opinion on any participant's position, and participants are told so. Coders work under confidentiality terms, documents never leave a controlled environment, and personal data inside contracts is redacted before coding where it is not needed.

  • Ratings: blocking means the gap stops licensing these records as scoped; fixable means a documented path exists, such as excluding a customer segment, amending a template or obtaining consent; not material means the code applies but does not affect the proposed use; undetermined means the evidence supplied is insufficient and is reported as its own category.
  • Coding is per dataset, and one dataset can carry several codes. The analysis reports both any-blocking-gap and the full code profile.
The evidence codebook: ten gap types and how each is rated
CodeGapEvidence that triggers the codeRated blocking when
G01Ownership unclearRecords or content created by contractors or vendors with no assignment or license backNo document shows the company holds or can grant the rights
G02Customer contract restrictionConfidentiality or use clauses limiting customer data to providing the serviceClause covers the records and no amendment path exists
G03Platform terms restrictionSoftware or platform terms limiting export or downstream use of hosted dataTerms bar the use and the data cannot be re-sourced
G04Privacy purpose mismatchNotice in force at collection did not cover disclosure for this purposeCounsel rule says compatibility cannot be shown and consent is impractical
G05Silent policy changeNotice or terms changed after collection to permit the use, without fresh notice or consentChange was retroactive and unannounced
G06Workforce contentRecordings, messages or monitoring data about employees with no matching policyNo policy or agreement covers the use
G07Embedded third-party materialAttachments, licensed content or partner documents inside the recordsMaterial cannot be removed without destroying the dataset's use
G08Sector confidentiality ruleRecords fall under a sector-specific confidentiality or disclosure lawCounsel rule says no permitted path exists
G09Prior grant conflictEarlier exclusive license, data-sharing deal or vendor right covering the same recordsExclusivity or field of use overlaps the proposed use
G10Unprovable permissionPermission likely existed but the company cannot produce consent records or signed termsProof cannot be reconstructed for the sampled period

Why are two of the codes there, and what evidence backs them?

G04 and G05 come straight from how regulators frame repurposing. Under GDPR Article 6(4), when personal data is processed for a purpose other than the one it was collected for and the processing is not based on consent or law, the controller assesses compatibility by weighing factors such as the link between purposes, the context of collection, the nature of the data, possible consequences and safeguards. The codebook does not perform that assessment; it records whether the evidence a counsel rule needs is present.

G05 reflects a U.S. concern. In a February 2024 Tech@FTC post, FTC staff wrote that a company that quietly changes its terms of service or privacy policy to allow more permissive data practices, such as using data for AI training, could be engaging in unfair or deceptive conduct. A study that coded only whether today's terms allow licensing would miss this gap, which is why coders must collect the version of each notice that was in force when the records were created.

The emphasis on documents over self-report has a precedent in AI data itself. The Data Provenance Initiative's audit of more than 1,800 text datasets reported license omission rates above 70 percent and error rates above 50 percent on widely used hosting sites. If dataset licenses are that often missing or wrong where they are supposed to be visible, a respondent's memory of what a customer contract says is a weak instrument.

How large a sample, which comparison and what kind of interval?

A defensible result is a weighted share of candidate datasets with a blocking gap, with a confidence interval that accounts for clustering of datasets within companies, from a probability sample whose outcome rates and nonresponse analysis are published. The sample size is set from the precision needed for the smallest subgroup the report will describe, with the design effect from clustering and weighting stated before fieldwork.

Response rates follow AAPOR's standard definitions so they can be compared with other surveys. Because the document request is burdensome, the study expects lower completion for documents than for interviews and treats documents as a second-phase sample. OMB's standards for federal statistical surveys call for a nonresponse bias analysis when the expected unit response rate is below 80 percent; this study adopts that threshold as its own trigger and compares interview-only and document-supplying companies on every frame variable.

The baseline is internal. The study compares the code profile of datasets that progressed to a signed agreement with those that stopped, which shows which gaps actually separate the two groups rather than which gaps are merely common. It also compares the respondent's stated stop reason with the document coding, which measures how often companies misdiagnose their own blockers.

What population does the study cover, and what conclusion would its design not support?

It covers candidate datasets at companies in the frame that considered licensing during the look-back period. It says nothing about companies that never considered licensing, and it cannot estimate how many companies could license data if they tried. It also cannot rank sectors on legal risk beyond the strata it samples.

It would not support any statement about whether a specific dataset may lawfully be licensed, any figure about deal values, or any claim about buyer demand. A blocking rating is a coded classification under published rules at one point in time, not a legal determination, and gaps rated fixable may still be expensive to fix.

Pre-registration and disclosure checklist

File the protocol with a time stamp before the first invitation, and publish with the eventual report every element AAPOR's disclosure standards ask for, including the sponsor and the party that conducted the work, the exact question wording, the sample design and frame, and the dates of data collection.

At a glance

Pre-registration and disclosure checklist

Checklist · 10 items

  1. Research question and the single primary measure: share of candidate datasets with at least one blocking gap.
  2. Population definition, look-back period, countries, sectors and employee floor.
  3. Frame source, known coverage gaps and the stratum targets.
  4. Screener wording and the rule for selecting up to three datasets per company.
  5. Interview instrument and document request, with version numbers.
  6. Codebook version, counsel sign-off date and the decision rules for G04 and G08.
  7. Agreement statistic, minimum threshold per code and the adjudication rule.
  8. Weighting method, variance estimation for clustered datasets and minimum cell size for publication.
  9. Nonresponse bias analysis plan and the response-rate formula used.
  10. Data handling: confidentiality terms, redaction, retention period and destruction date for collected documents.

0 of 10 ticked. Tick each item as you go.

Using the codebook on your own candidate datasets

A company does not need the study to use the codebook. Run the ten codes against one candidate dataset, collect the notice versions that were in force when the records were created, and rate each code blocking, fixable, not material or undetermined. The undetermined pile is usually the to-do list: documents nobody has pulled yet.

For a company that decides to license selected records directly to SourceX, SourceX runs rights review alongside sourcing, contracting, preparation, delivery and payment. A participating company signs off on scope and terms before a single record leaves its hands. The first step is a brief company assessment that does not require uploading any data.

Frequently asked questions

Has this rights gap study been run?

No. This page publishes the protocol, instruments and codebook only, and it reports no findings. Any published share of deals stopped by rights gaps should come with its own frame, dates, denominator and method; without them it is an estimate of unknown quality.

What does the survey and document coding protocol consist of?

A screener, an interview instrument, a document request, a versioned codebook with rating rules, a coder training and adjudication process, an agreement threshold, a weighting and variance plan, and a disclosure plan. Counsel signs off the codebook and the decision rules, not individual companies' positions.

Why sample datasets rather than companies?

Rights positions differ between record families inside one company. Support tickets may be governed by customer contracts while engineering tickets are not, so a company-level answer would average away the gaps the study is trying to find.

Can survey answers alone measure rights gaps?

Not reliably. Respondents often know whether a deal stopped but not which clause stopped it, and stated reasons tend to name the last obstacle rather than the first. Coding the governing documents and comparing them with the stated reason is what makes the measure checkable.

What would make the result defensible?

A probability sample with published outcome rates, a nonresponse bias analysis, intervals that account for several datasets per company, agreement statistics for every code, and a comparison between datasets that reached signature and those that stopped.

Sources

  • AAPOR's Standard Definitions (10th edition, 2023) standardize final disposition codes for sampled cases and the formulas for response and other outcome rates. Source
  • AAPOR published Standard Definitions for Establishment Surveys in December 2025, extending disposition codes to business and establishment data collections. Source
  • AAPOR's disclosure standards ask researchers to name the sponsor and conductor, give exact question wording, describe the sample design and frame, and state the dates of data collection. Source
  • OMB's 2006 Standards and Guidelines for Statistical Surveys tell federal agencies to plan a nonresponse bias analysis when the expected unit response rate is below 80 percent. Source
  • The Census Bureau's Statistics of U.S. Businesses reports counts of firms, establishments, employment and payroll by industry and enterprise employment size. Source
  • GDPR Article 6(4) lists factors for assessing whether processing for a new purpose is compatible with the original one, including the link between purposes, the context of collection, the nature of the data, possible consequences and safeguards. Source
  • In a February 2024 Tech@FTC post, FTC staff wrote that quietly changing terms of service or privacy policies to permit practices such as AI training could be unfair or deceptive. Source
  • The Data Provenance Initiative audited more than 1,800 text datasets and reported license omission above 70 percent and error rates above 50 percent on widely used hosting sites. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies