Deal economics
How to diagnose why a buyer rejected your business dataset
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Diagnose a rejected business dataset by sorting the buyer's reason into value, rights, privacy, quality or fit, confirming it with specific evidence, and matching it to one remediation. Price is often a cover for lower usefulness, unclear rights or a failed sample. Interview the buyer within a week, code the answers independently, then fix and re-offer, narrow the scope, or stop.
Value drivers
Dataset valuation buyer
Key takeaways
- Treat the first reason a buyer gives as a hypothesis; price often stands in for a fit, rights or quality problem.
- Sort every decline into value, rights, privacy, quality, fit or timing, and keep privacy separate from rights.
- Ask a reversal question: which single change would have moved the decision to a yes.
- Code answers with two independent coders and an evidence grade before spending on any fix.
- Use the remediation matrix to choose between re-offering, narrowing the scope, waiting or stopping.
How do you diagnose why a buyer rejected a business dataset?
Diagnose a dataset rejection by sorting the buyer's reason into one of five families (value, rights, privacy, quality or fit), confirming it with evidence rather than the buyer's first sentence, and then matching it to one remediation. A decline is information about the offer, not a verdict on the records. The goal is to learn whether a change you control would have changed the answer.
The decision rule is simple. If the root cause is something you can fix with documentation or preparation, fix it and re-offer to the same kind of buyer. If it is something only scope can fix, narrow the offer. If it is something nothing you control can fix, such as a buyer whose model goal never matched your records, record it and stop spending time on that buyer type.
This page gives you the two tools that make the rule usable: a short interview to run after a decline and a remediation matrix that turns each coded reason into a next step. It deliberately does not attach prices or amounts to any reason. Value in this context is a conditional range that moves with rights, coverage and buyer purpose, so the useful question is which condition moved, not what number was lost.
What do the five rejection families, plus a timing decline, really mean?
Nearly every rejection reason you will hear fits one of five families, with timing declines as a sixth case that says nothing about the data, and each points to a different owner inside your company. Keep privacy separate from rights: being allowed to access records, having privacy permission to share them, owning them and holding contractual rights to license them are four different questions, and a buyer can decline on any one of them.
Use the table to translate what a buyer says into what you should test first.
| Family | What the buyer usually means | Evidence to ask for | Owner on your side |
|---|---|---|---|
| Value (price) | Expected usefulness for their model goal is lower than the cost and effort of licensing and preparing the records | Which use case they priced against, and which gap lowered expected usefulness | Commercial lead |
| Rights | They cannot show a clean chain from your company's contracts and policies to the uses they need | The specific contract, policy or third-party content they flagged | Counsel and records owner |
| Privacy | Personal or sensitive information remains, or the de-identification method is not documented to their standard | The record types and fields involved, and the method they expected | Privacy lead |
| Quality | A sample showed errors, gaps, duplicates or labels they cannot trust | The sample IDs and the checks that failed | Data or operations lead |
| Fit | The records do not cover the task, period, language, segment or volume their project needs | The spec they compared you against | Commercial lead |
| Timing or process | Their budget, project or review window moved; nothing about the data changed | Whether and when the project resumes | Commercial lead |
Why is the stated reason often not the real one?
The first reason a buyer gives is usually the least costly one to say, so treat it as a hypothesis. Price is the most common cover reason, because saying the offer costs too much is easier than explaining that rights look uncertain or that a sample failed internal review.
Three patterns recur. A value objection often hides a fit gap: the buyer needed outcomes or a longer history, so the same records were worth less to them. A quality objection often hides a rights or privacy concern: heavy redaction applied to reduce risk can strip the content a buyer wanted. A fit objection sometimes hides a documentation gap: the buyer could not tell from your description whether the records cover their task, so they assumed they did not.
Documentation matters because buyers can only evaluate what they can see. Datasheets for Datasets proposed that every dataset carry a record of its motivation, composition, collection process and recommended uses, and a decline after a thin description is often a decline of the description. Rights documentation has the same problem at scale: the Data Provenance Initiative audit of widely used AI datasets reported license omission above 70 percent and license error rates above 50 percent on popular hosting sites, which is why many buyers now ask for rights evidence before they ask about volume.
The rejection reason interview
Run a short interview within a week of the decline, while the reviewer still remembers the sample, and ask for evidence rather than opinions. Keep it to 20 minutes, send the questions in advance and promise that the answers are used only to improve the offer. Ask every question, even when the buyer's email already gave a reason.
At a glance
The rejection reason interview
The questions below are the interview half of the asset. Record each answer verbatim, then code it with the matrix in the next sections.
- 1. Purpose: Which model, evaluation or product goal did you assess the dataset against?
- 2. First blocker: At what point in your review did the dataset stop moving forward, and who stopped it (technical, legal, privacy, procurement)?
- 3. Value: Which part of the dataset would have been useful, and which part would you not have used at all?
- 4. Rights: Was there a specific contract, policy, third-party content type or permitted-use term you could not get comfortable with?
- 5. Privacy: Which fields or record types raised a privacy concern, and what de-identification method or documentation did you expect?
- 6. Quality: Which sample records or checks failed, and how many did you review?
- 7. Fit: Which requirement did the dataset miss: task coverage, period, volume, language, segment, outcomes or format?
- 8. Alternatives: What are you using instead, in general terms (another source, synthetic data, your own collection, nothing yet)?
- 9. Reversal test: If one thing changed, which change would have moved this to a yes?
- 10. Second blocker: If that change had been made, what would most likely have stopped it next?
- 11. Timing: Is the project paused, canceled or proceeding with another source?
- 12. Permission: May we come back if we fix the issue you named, and who should we contact?
How should you code the answers so two people agree?
Code each interview with one primary reason, at most two secondary reasons and an evidence grade, and have two people code it independently before comparing. Independent coding stops the person who ran the sale from recording the reason that is easiest to accept.
Use three evidence grades. Grade A means the buyer named specific records, fields or terms. Grade B means the buyer named a category but no specifics. Grade C means the reason is inferred from silence or a one-line email. Only grade A and B reasons should drive remediation spending; grade C reasons go on a watch list.
Once you have enough coded declines to compare, measure how often your two coders agree on the primary family. Cohen's kappa is the standard statistic for agreement between two human coders; the scikit-learn documentation describes it as intended to compare labelings by different human annotators and notes that scores above 0.8 are generally considered good agreement. With only a handful of declines, skip the statistic and resolve each disagreement in a short adjudication note. Disagreement is itself useful: it usually means the buyer's answer mixed two families, and the reversal question (question 9) settles which one came first.
- Primary family: value, rights, privacy, quality, fit or timing.
- Reason code: the specific row from the remediation matrix.
- Evidence grade: A (specific), B (category), C (inferred).
- Reversal answer: the single change the buyer said would move the decision.
- Next blocker: what the buyer said would have stopped it after that.
- Coder initials, date and adjudication note if the coders disagreed.
The remediation matrix
The remediation matrix maps each coded reason to a diagnostic test you can run yourself, the remediation that addresses it and whether the outcome is to re-offer, narrow or stop. Run the diagnostic before you spend on remediation, because a reason that does not survive its own test was a cover reason.
At a glance
The remediation matrix
Use it as is, adding rows only when a new reason appears in at least two coded declines.
| Reason code | Diagnostic test you can run | Remediation | Typical outcome |
|---|---|---|---|
| V1 Expected usefulness too low | Map the buyer's stated goal to the fields that serve it; count records that carry them | Re-scope the offer to the records that serve the goal; document what the rest can and cannot support | Re-offer narrowed |
| R1 Contract restriction | List customer and vendor contracts covering the records; flag confidentiality and use clauses | Counsel review; exclude restricted record families or obtain consent where possible | Narrow or stop for that family |
| R2 Third-party content | Sample for embedded third-party documents, images or code | Remove or separately clear third-party content; label what stays | Re-offer after removal |
| P1 Residual personal data | Re-run the redaction check on the buyer's flagged sample | Fix the method, rerun on the full set, document the method used | Re-offer with method note |
| P2 Method not documented | Check whether your package explains how personal data was handled | Write the method, scope and residual-risk statement; for health records follow the HHS guidance and counsel | Re-offer with documentation |
| Q1 Errors or unreliable labels | Re-review a random sample of the flagged field against source evidence | Correct, relabel or flag unreliable values; report the measured error rate | Re-offer with quality report |
| Q2 Gaps or duplicates | Measure completeness and duplicate share by period and record type | Deduplicate, mark gaps, narrow to complete periods | Re-offer narrowed |
| F1 Coverage mismatch | Compare the buyer's spec line by line with your inventory | None if the records do not exist; document coverage for the next buyer | Stop for that buyer type |
| T1 Timing | Confirm whether the project resumes | None; diary a follow-up | Wait |
Which assumption decides whether a re-offer can succeed?
The assumption with the largest effect on buyer fit is usually the permitted use, because it decides which buyers can use the records at all before anyone looks at quality. A dataset that is clean, complete and well documented still fails if the uses a buyer needs fall outside what your contracts and policies support.
The second largest is coverage of the buyer's task. The EU AI Act states the general principle well for high-risk systems: training, validation and testing data sets should be relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose. The phrase that matters for a supplier is in view of the intended purpose. The same records can fit one buyer and fail another, so a fit decline is a statement about a purpose, not about the records.
Show uncertainty as conditions rather than numbers. Instead of a single estimate of value, write the offer as scenarios: what the records support if outcomes are included, if only a narrower period is cleared, or if a record family is excluded for rights reasons. Each scenario names the assumption it depends on, so the next decline tells you which assumption failed. Quality claims need the same discipline. Even curated benchmark test sets carry label errors, estimated at an average of at least 3.3 percent across ten widely used test sets in one study, so report a measured error rate with its sample size instead of describing records as clean.
Illustrative worked scenario: three declines, one root cause
Illustrative example, invented to show the method: a regional HVAC service company with several years of job records in a field service system offered work orders, technician notes and invoices to AI developers building service agents. Three buyers declined within two months.
Buyer one said the offer was too expensive. The interview showed the buyer had assessed the records against a diagnosis task and could not find the final fix in most jobs, because outcomes sat in invoice line items that were not linked to the work order. Coded V1, secondary Q2, grade A. Buyer two said the records were not a fit. The reversal answer was that linked outcomes would have moved it to a yes. Coded F1 on the first pass, recoded to Q2 after adjudication because the outcomes existed and only the link was missing. Buyer three cited privacy: customer names and addresses remained in free-text notes. Coded P1, grade A.
Read together, two of three declines traced to one root cause, unlinked outcomes, which the company could fix by joining invoices to work orders on the job number. The privacy issue was a separate preparation fix. The company linked outcomes, reran redaction on notes, wrote a one-page method note and re-offered a narrowed set covering the years where the link held. The example shows the method only; it does not describe a real company or predict how any buyer would respond.
When is a decline a reason to stop offering the records?
Stop when the coded reasons show a fit gap you cannot close with records you have, or when the same rights reason appears twice and counsel confirms it cannot be cured. Spending on preparation does not fix a purpose mismatch or a contract that forbids the use.
At a glance
When is a decline a reason to stop offering the records
Narrow the offer when the problem is concentrated: one record family with rights restrictions, one period with poor data, one segment with personal data you cannot remove without destroying the content. Narrowing almost always beats a broad offer with caveats, because buyers review the weakest part first.
Bring in counsel and a privacy lead before re-offering whenever a decline involves contracts, consent, health, financial or employee records, or records about people in more than one jurisdiction. Removing identifiers can answer a privacy objection, yet it never supplies a licensing right that a contract or policy withholds. For health information covered by HIPAA, HHS describes two de-identification methods, Expert Determination and Safe Harbor, and which one fits is a question for counsel and a qualified expert, not for the sales follow-up.
- Stop: fit gap with no matching records, or an incurable rights restriction confirmed by counsel.
- Narrow: restricted, poor or sensitive data concentrated in identifiable families, periods or segments.
- Fix and re-offer: documentation, linkage, redaction method or quality reporting gaps.
- Wait: timing declines with no data-related reason.
Where SourceX fits after a decline
After a decline, the hard part is separating what can be cured from what cannot. Licensing selected records to AI developers through SourceX runs through sourcing and qualification, rights review, contracting, data preparation, delivery and payment, so curable and incurable reasons surface inside one process. Because rights are reviewed before preparation starts, a rights problem surfaces before anyone spends effort cleaning records that could not be licensed anyway.
Volume is a common hidden fit reason. Measured in volume, the usual buyer expectation in SourceX's general guide is two or more years of consistent history with tens of thousands of records in one system, and rarer work can sometimes qualify below that; each case is judged on its own. Nothing leaves your company until it has approved the scope and terms and signed the agreement.
Checking whether a declined dataset deserves a second attempt starts with a brief assessment of your company, with no data upload required.
Frequently asked questions
What should the rejection reason interview and remediation matrix include?
The interview should cover the buyer's purpose, the first blocker and who raised it, the value, rights, privacy, quality and fit questions, a reversal question and the next likely blocker. The matrix maps each coded reason to a diagnostic test, a remediation and an outcome of re-offer, narrow, stop or wait.
Is price ever the real reason a buyer rejects a dataset?
Sometimes, but a price objection usually reflects lower expected usefulness for the buyer's specific goal. Ask which use case the buyer assessed against and which part of the data they would not have used. If the answer names a gap in coverage or outcomes, the issue is fit or quality, not price.
How do I get a buyer to explain a rejection honestly?
Ask soon after the decline, send short questions in advance and ask for specifics such as sample IDs, fields or contract terms rather than opinions. The reversal question, which single change would have moved the decision, usually produces the most honest answer.
Should I fix every reason a buyer gives?
No. Run the diagnostic test first and spend only on reasons supported by specific evidence. Fit gaps with no matching records and incurable rights restrictions are reasons to stop or narrow, not to invest in preparation.
Does de-identifying the records solve a rights rejection?
No. Privacy permission, ownership and contractual licensing rights are separate questions. De-identification can address a privacy concern, but it does not create the right to license records that a contract or policy restricts; counsel reviews those facts.
How many declines do I need before patterns mean something?
There is no fixed number. Two declines with the same coded root cause and grade A evidence are usually enough to act on. A statistic such as Cohen's kappa for coder agreement only becomes useful once you have enough coded declines to compare.
Sources
- Datasheets for Datasets proposes that each dataset document its motivation, composition, collection process and recommended uses. Source
- The Data Provenance Initiative audit reports license omission above 70 percent and error rates above 50 percent on widely used dataset hosting sites. Source
- The scikit-learn documentation describes Cohen's kappa as a measure for comparing labelings by different human annotators, with scores above 0.8 generally considered good agreement. Source
- Article 10(3) of the EU AI Act requires training, validation and testing data for high-risk systems to be relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose. Source
- Northcutt, Athalye and Mueller estimate an average of at least 3.3 percent label errors across the test sets of 10 widely used datasets. Source
- HHS guidance describes two methods of de-identifying protected health information under the HIPAA Privacy Rule: Expert Determination and Safe Harbor. Source
See if your company qualifies
A short company assessment. No data uploads are needed.