Leadership and readiness
Data licensing owner concerns survey: a protocol for interviewing owners
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
This page is a study protocol and reports no results. To find which questions business owners need answered before a data deal, interview decision-makers from a disclosed, stratified sample, ask unprompted questions before showing any topic list, code each question by concern family and by whether the owner could get it answered, double-code a share of transcripts, and report counts only with their denominators and recruitment method.
License terms
Data licensing owner
Key takeaways
- This is a method for a future study; no interviews have been run and no findings are reported here.
- The core measure is an unanswered question: one the owner says must be resolved before agreeing, and that the information available to them did not resolve.
- Unprompted questions come first and are coded separately from answers to a topic card, so the guide does not plant the concerns it measures.
- A codebook with definitions and exclusions, double coding and a chance-corrected agreement statistic make the coding auditable.
- A purposive interview sample can show which questions exist and how they cluster; it cannot estimate how common they are among all owners.
How could research identify the questions owners need answered before a data deal?
Interview the people who would decide, before showing them any list of topics, and record each question they say would have to be answered before they agreed to license company data. Then code those questions with a published codebook, mark which ones the owner could already answer, and report the patterns with the sample, recruitment method and denominators in full.
This page is the protocol for that data licensing owner concerns survey: the research question, population and sampling, the neutral interview guide, the coding plan, reliability and analysis rules, ethics and disclosure, and a pre-registration checklist. It contains no results, no counts and no quotations from owners. Any example below is labeled illustrative.
The research question is: among decision-makers at operating companies who have considered or been asked about licensing business records for AI, what questions do they need answered before agreeing, which of those remain unanswered after the information they had, and who did they expect to answer them?
What exactly is an unanswered question?
An unanswered question is one the participant says must be resolved before they would agree, and which the information available to them at the time did not resolve. Defining it this way separates three things that often get blurred in surveys: a general worry, a question, and a blocker.
Each mention is coded on four attributes in addition to its topic: whether it came up unprompted or after the topic card; its resolution status (answered, partly answered, unanswered, or answered but not believed); its weight (would block the deal, would change terms, or would only slow it down); and who the participant expected to answer it, such as their counsel, their accountant, the counterparty, a regulator or nobody they could name. The last attribute is often the most useful, because a question with no expected answerer is a gap in the market, not just a worry.
Who should be interviewed, and how are they found?
Interview the person who would sign or veto: an owner, chief executive, or the finance or operations executive with that authority. Stratify by the characteristics most likely to change the questions, and disclose every recruitment channel, because each one brings its own bias.
There is no public list of owners who have considered data licensing. The Census Bureau's Annual Business Survey samples nonfarm employer firms and collects owner characteristics, which makes it a useful reference for describing how a study sample differs from employer businesses in general, though not a contact list. Nonemployer businesses are outside that survey, so a study that includes them needs a separate frame.
| Stratum | Levels | Why it may change the questions | Recruitment note |
|---|---|---|---|
| Exposure to a data deal | Never approached; approached and declined; in discussion; signed | Questions shift from whether to how once a deal is concrete | People who signed are easiest to find and least typical |
| Record type | Customer communications; transactions; operational logs; documents and files | Privacy and confidentiality questions depend on what the records contain | Classify by the records discussed, not the industry label |
| Ownership | Founder-led; family; private equity backed; subsidiary | Investor or parent consent and succession raise distinct questions | Ask who else must approve |
| Company size | Bands by employee count | Larger firms route questions to counsel and procurement | Record the band, not exact headcount |
| Regulatory exposure | Sector rules apply or do not | Regulated records produce compliance questions first | Do not ask for regulated data itself |
| Recruitment channel | Industry association; professional networks; inbound to a data company; journalist contacts | Inbound contacts are already interested and self-selected | Report results by channel |
How many interviews, and when do you stop?
Set a planned minimum per stratum and a stopping rule in advance, then report where saturation was and was not reached. Saturation depends on what you need: in Hennink, Kaiser and Marconi's study of 25 interviews, code saturation, where the range of issues was identified, came at nine interviews, while meaning saturation, a full understanding of those issues, needed 16 to 24. A study that wants to explain why owners hold a question, not only list it, should plan for the second.
Use a stopping rule tied to the codebook: stop recruiting in a stratum when a pre-set number of consecutive interviews adds no new code and no new resolution pattern, and only after the planned minimum. Record the interview at which each code first appeared, so readers can judge saturation themselves.
The neutral interview guide
The guide moves from open to specific, and the interviewer never names a concern before the participant does. Neutrality rules: no figures for payouts or prices, no named buyers or platforms, no description of any company's process, and the same wording and order for every participant. Questions are numbered so coders can locate answers.
At a glance
The neutral interview guide
- Part A, unprompted. A1: In your own words, what would licensing some of your company's records for AI training involve? A2: If someone proposed it next month, what would you need to know before saying yes? A3: Which of those could you answer today, and how? A4: Who would you ask about the rest? A5: Is there anything that would make you say no regardless of the answers?
- Part B, prior experience. B1: Has anyone approached you about this, or have you looked into it? B2: What happened next, and at what point did it stop or continue? B3: Which question took longest to answer, and was it answered in the end?
- Part C, topic card. Show a card listing broad topics in randomized order: rights and ownership, customer privacy, confidentiality, value, effort, reputation, control over use, security and delivery, contract terms, ending the arrangement. C1: Which of these, if any, did you not mention but would matter? C2: For each one you pick, what is the question you would need answered?
- Part D, resolution. D1: For the questions you raised, what kind of answer would be enough, such as a document, a contract clause, advice from your own counsel or someone else's experience? D2: Is there a question you think nobody could answer for you?
- Close. E1: Is there anything we should have asked? Thank the participant and confirm how their words may be used.
The coding plan
Code the transcript, not the interviewer's notes, at the level of one question or concern per mention. Each mention gets one concern family from the codebook and the four attributes defined above. Families are deliberately broad; subcodes are added only through a logged codebook revision.
- An eleventh family, Other, holds mentions that fit none of the above; any family that collects a recurring theme in Other is proposed as a new code and logged.
- Illustrative coding: a participant asks, unprompted, whether a customer contract signed years ago lets the company reuse support emails, and says their counsel would need to check. Code: Rights and ownership; unprompted; unanswered; would block; expected answerer counsel.
| Concern family | Definition | Include | Exclude |
|---|---|---|---|
| Rights and ownership | Whether the company may license the records at all | Customer contracts, vendor terms, investor or parent consent | Privacy law questions, coded separately |
| Privacy of people in the records | Obligations to customers, employees and others named | Consent, notice, de-identification, deletion requests | Commercial confidentiality |
| Confidentiality and competition | Exposure of business information | Pricing, suppliers, strategy, trade secrets | Personal data |
| Value and payment | How value is set and paid | Basis for valuation, timing, tax treatment | Effort or cost of preparation |
| Effort and disruption | Work required from staff and systems | Export, review time, leadership attention | Security of transfer |
| Reputation | How customers, staff or press would react | Disclosure, public perception, employee reaction | Legal liability |
| Control over use | What the data may be used for and by whom | Permitted uses, sublicensing, competitors, model outputs | Termination |
| Security and delivery | How data leaves and is protected | Transfer method, storage, access, breach | Privacy law duties |
| Contract and liability | Terms, warranties and remedies | Indemnities, audits, governing law | Valuation |
| Reversibility and exit | Whether the decision can be undone | Termination, deletion, exclusivity, renewal | Initial permitted uses |
Double coding, agreement and what counts as a finding
Two coders independently code a pre-set share of transcripts, chosen at random within each stratum, before the codebook is frozen for the main analysis. With two coders on the same transcripts, Cohen's kappa is the usual agreement statistic; it adjusts for agreement expected by chance, which percent agreement does not, so report both for the concern family and for each attribute.
Disagreements are resolved by discussion and logged with the codebook change they caused. A third reviewer with experience of data or technology contracts audits a sample of the block-weight codes, because the boundary between a blocker and a term to negotiate is where coders most often differ.
A finding is a pattern reported with its denominator and its channel: how many participants in which strata raised a family, unprompted or prompted, and how many left it unanswered. Do not convert counts into percentages for all owners, and do not compute confidence intervals as if the sample were random; it is not. Quotations are used only with consent, lightly edited for identity and never chosen to fit a narrative.
Disclosure, consent and the sponsor's conflict of interest
Publish who paid for the study and who ran it, and keep the sponsor out of recruitment decisions and coding. A data licensing company has an obvious interest in which concerns appear and how they are framed. AAPOR's disclosure standards ask researchers to name the sponsor and the party that conducted the research, to describe how participants were contacted, selected or recruited along with eligibility rules, and to publish the exact wording of measurement tools, which includes interview scripts.
For reporting, follow COREQ, the 32-item checklist for reporting interview and focus group research, which covers the research team, study design and analysis. For ethics, consent covers recording, storage, quotation and withdrawal. If the study runs through an institution subject to the U.S. Common Rule, research limited to interview procedures may be exempt under 45 CFR 46.104(d)(2) when it meets the listed conditions, one of which concerns whether disclosure of responses could reasonably place participants at risk of liability or damage to their financial standing, employability or reputation. An owner discussing customer data or contract terms may well meet that risk test, so an institutional review decides, not the research team.
Do not collect business records, customer data or contract documents in the interview. The study is about questions, not about any company's data.
Pre-registration checklist and the conclusions this design cannot support
Register the plan, time-stamped, before the first interview. The Center for Open Science describes preregistration as a record of pre-specified plans made before seeing results, and expects departures from the plan to be disclosed rather than hidden. The registration should include the items below.
At a glance
Pre-registration checklist and the conclusions this design cannot support
Checklist · 9 items
- Research question and the definition of an unanswered question.
- Population, strata, recruitment channels and eligibility rules.
- Planned minimum per stratum and the saturation stopping rule.
- Interview guide version, card randomization and interviewer training.
- Codebook version, attribute definitions and revision procedure.
- Double-coding share, agreement statistics and adjudication process.
- Reporting rules: denominators, results by channel, quotation consent.
- Sponsor, conductor and the sponsor's role, if any, in each step.
- Data retention, de-identification of transcripts and deletion dates.
0 of 9 ticked. Tick each item as you go.
Where a protocol like this meets an actual decision
The design cannot support claims about how common a concern is among all business owners, whether a concern causes deals to fail, or how owners in unsampled sectors or countries think. It can show which questions exist, how they cluster, which stay unanswered and whom owners expect to answer them, which is what an owner, journalist or adviser needs to plan the next conversation.
An owner with live questions does not need to wait for a study. SourceX's first step is a short company assessment that needs no data uploads, and the company approves the scope and terms and signs before any data is delivered, so the questions in this guide can be worked through with counsel before anything leaves the company.
Which questions measure business data supplier motivations without leading the owner?
Add a Part F, asked after Part D and before the close, that covers four motivations in a fixed order: compensation, workload, trust and restrictions. Each question asks what would have to be true, never how much, so the guide's no-figures rule holds and every answer can later be checked against what the company actually did.
Tversky and Kahneman's 1974 paper in Science describes adjustment from an anchor as one of three heuristics people use for numerical judgments under uncertainty, and a figure spoken by an interviewer, or read in a recent headline, can become that anchor. Ask about form, timing and comparison instead, and record any number the owner volunteers as volunteered, with its stated origin.
Owners can use the five questions to prepare for a first call with any licensing partner: written answers state, in their own terms, what a partner must meet and which uses stay off the table.
- F1 Compensation: What would the company need to receive for this to be worth doing, in what form and on what timing? What are you comparing it with?
- F2 Workload: Who would do the export and review work, and how many of their hours could you spare before saying no? What would they stop doing to make room?
- F3 Trust: What would you need to see, read or hear, and from whom, before believing the other party will use the records only as agreed?
- F4 Restrictions: Which uses, recipients or terms would you rule out whatever the compensation? Which would you accept only with a condition?
- F5 Trade-off: If only one of these four could change, which would most change your answer?
- Coding: each Part F answer is coded driver, condition or veto, and kept apart from the concern codes, so a veto on restrictions is never averaged with a driver on compensation.
How do stated motivations line up with what a supplier cohort actually does?
Link each interview to the same company's later path, stage by stage, and compare what the owner said with how far the company got. A motive that predicts nothing about who reaches acceptance is a talking point, not evidence.
Define a cohort by the month a company first applied, and date every stage. Log supplier-side hours and compare them with the owner's own F2 estimate, because preparation that runs far past that estimate is a plausible reason to withdraw. Companies still inside a stage are not failures: survival analysis calls them right-censored, and the lifelines documentation shows that ignoring them, or counting their current durations as complete, underestimates the true average duration.
Capacity follows from Little's law, L = lambda W, which John Little revisited in Operations Research in 2011: the average number in a stable system equals the arrival rate times the average time spent there, so companies in preparation equal the signature rate times the average preparation time. If preparation time rises with signatures, reviewers saturate and acceptance stalls, however many owners say yes.
| Stage | Event that moves a company | Denominator | Motivation evidence attached | Effort logged |
|---|---|---|---|---|
| Interviewed | Completed Parts A to F | Companies invited in the cohort month | Codes for F1 to F5 | Interview minutes |
| Qualified | Passed eligibility and dataset fit | Interviewed | Whether a veto code reappears as the stated reason for any exit | Reviewer hours on qualification |
| Rights cleared | Review found the offered records licensable | Qualified | Restriction codes compared with the terms requested | Supplier counsel hours, if disclosed |
| Signed | Company approved scope and terms and signed | Rights cleared | Compensation and trust codes compared with the final terms | Negotiation rounds |
| Prepared | Export and de-identification complete | Signed | Workload code compared with actual supplier hours | Supplier staff hours against the F2 estimate |
| Accepted | Buyer review accepted the dataset | Prepared | None new: this is the outcome being explained | Rework cycles |
Do more applications mean more accepted supply?
Not necessarily; the cohort model tells which. More applications can add screening and review work while accepted supply stays flat, so compare accepted datasets per qualified company and per reviewer hour across cohorts, never application counts alone. For owners this matters: a program judged on applications has a reason to recruit widely; one judged on accepted datasets has a reason to screen early and say when records will not fit.
Illustrative counts, not SourceX data: cohort A has 40 applications, 10 qualified and 4 accepted; cohort B, after wider outreach, has 120 applications, 15 qualified and 4 accepted. Applications tripled; accepted supply did not move. If cohort B's interviews show more compensation drivers, and its supplier hours run far above the owners' own F2 estimates, the campaign reached owners drawn by price who stopped at the preparation burden: a targeting finding, not a demand finding.
Count the case that challenges any headline: an owner who names compensation as the main driver, then exits at rights clearance over a restriction first raised on seeing the terms. Report such cases by stage, code and denominator, not as anecdotes.
What could make the motivation evidence wrong, and what should an investor ask?
Four things could mislead. Stated motives may not predict behavior, which is why the cohort link exists. Small cohorts cannot support rates, so every count carries its denominator. Channel mix changes who enters a cohort, so results stay split by recruitment channel, as the protocol requires. And a change in privacy rules or buyer demand can move every stage at once: a market effect, not a motivation effect.
An investor weighing any supplier program, SourceX included, can ask four things: each cohort month's denominator at every stage, and how companies still in progress are counted; which motivation codes separate companies that reached acceptance from those that left, and where they left; how supplier hours compared with the workload owners predicted before signing; and whether accepted datasets per reviewer hour rose or fell as applications grew.
Methods note: the module, cohort table and capacity rule were written on October 26, 2026 from this page's protocol, Little's law and standard survival analysis. No interviews or cohorts have been measured and no results are published here; every count above is illustrative. SourceX shares company and commercial metrics with investors directly; investor materials are requested through its investor page.
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 for this section
- Tversky and Kahneman (Science, 1974) describe adjustment from an anchor as one of three heuristics people use in numerical judgments under uncertainty. Source
- The lifelines documentation defines right-censored cases as those whose event has not yet been observed and shows that ignoring them, or treating their current durations as complete, underestimates the true average duration. Source
- John Little's 2011 Operations Research article revisits the queueing formula L = lambda W, which relates the average number in a system to the arrival rate and the average time spent in it. Source
Frequently asked questions
Does this page report what business owners think about data licensing?
No. It is a protocol for a study that has not been run: an interview guide, codebook, sampling plan and disclosure rules. Any example is illustrative, and findings should be published only with the sample, recruitment channels and denominators.
What should the neutral interview guide and coding plan include?
Unprompted questions before any topic list, a randomized topic card, and questions on how each concern could be resolved and by whom. The coding plan needs concern families with definitions and exclusions, per-mention attributes such as resolution status and weight, double coding and a logged revision process.
Which owners would the study describe, and what could it never claim?
It covers decision-makers at operating companies who have considered or been asked about licensing records, sampled purposively across strata. It cannot estimate how common a concern is among all owners or show that a concern causes deals to fail.
How many interviews, and what reporting, would make the findings credible?
A planned minimum per stratum with a pre-registered saturation rule, results reported by recruitment channel, the interview at which each code first appeared, and chance-corrected agreement between coders. Counts should carry denominators and should not be presented as population percentages.
Why not use an online survey instead of interviews?
A closed survey has to list the concerns in advance, which measures recognition of someone else's list rather than the questions owners bring themselves. Interviews come first to build the codebook; a later survey can then test how common each coded question is in a probability sample.
Sources
- AAPOR asks researchers to disclose the sponsor and conductor, how participants were recruited and eligibility rules, and the exact wording of interview scripts. Source
- In 25 interviews, code saturation came at nine interviews and meaning saturation needed 16 to 24. Source
- COREQ is a 32-item checklist for reporting interviews and focus groups. Source
- The Common Rule exempts some research limited to interview or survey procedures when listed conditions are met. Source
- The Annual Business Survey samples nonfarm employer firms and collects owner characteristics; nonemployers are not surveyed. Source
- Cohen's kappa fits two raters in a fully crossed design and adjusts for chance agreement, unlike percent agreement. Source
- Preregistration records plans before results are seen, and departures from the plan should be disclosed. Source
See if your company qualifies
A short company assessment. No data uploads are needed.