Deal economics
Data rejection vs commercial terms: log two verdicts, not one
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Separate data rejection from commercial terms by giving every opportunity two verdicts: a data verdict (passed, failed or not tested against the agreed acceptance test) and a terms verdict (agreed, failed or not reached). A single rejected status hides which one ended the deal. Compute data failures only over tested deals and terms failures only over deals that reached terms, and treat early exits as competing outcomes, not missing data.
License terms
Data rejection
Key takeaways
- A data rejection is a failed acceptance test; a terms rejection is a failed negotiation, and a deal can end on either, both or neither.
- Give each opportunity two three-state verdicts: data passed, failed or not tested, and terms agreed, failed or not reached.
- Compute data failures over tested deals and terms failures over deals that reached terms; mixed denominators distort the funnel.
- In the illustrative 40-offer funnel, a 57.5 percent rejection rate splits into about 26 percent data failure and about 47 percent terms failure.
- Treat terms exits as competing outcomes, not censored data, and route authority problems to rights review rather than pricing.
Did the buyer reject the data, the terms, or neither?
A lost dataset deal can end on the data, on the commercial terms, on both, or before either was tested, and those are different events with different owners. A data rejection means the records failed an agreed acceptance test. A terms rejection means the parties could not agree on price, scope of use, duration, exclusivity or risk allocation, whatever the records were like.
Contract law already keeps the two apart. Under U.S. sales law, UCC section 2-601 lets a buyer reject a delivery that fails in any respect to conform to the contract, which presupposes that a contract exists. A negotiation that never reaches agreement produces no contract and nothing to reject. Data licenses are not always sales of goods, but the line is useful: rejection is a verdict on conformity, while a failed negotiation is a verdict on the deal.
For a delivery or acceptance owner, the decision is practical. What should each opportunity record, and how should rates be computed, so that the fix the team chooses matches the failure that actually ended the deal? The evidence that settles it is a dated test result for the data, the last written positions for the terms, and the order in which each happened.
Why one rejected status sends teams to fix the wrong thing
A single rejected status merges four outcomes into one number, so teams fix whatever they own rather than what broke. If terms are the real leak, the delivery team re-cleans records that were never the problem. If the data failed its test, the commercial team discounts a price that was never the problem.
The worst merge is the untested deal. When a negotiation collapses on price before anyone runs the acceptance test, the data verdict is unknown, yet a single status files it as rejected and the records take the blame. Over a quarter, those untested exits can make clean data look like a quality problem.
Rights add a second trap. A supplier that cannot grant training use looks like a party refusing terms, but the cause is authority, not preference, and no price fixes it. Buyers screen for this separately from quality: OpenAI's data partnerships page, for example, states that it is not after datasets that hold sensitive or personal information, or that hold information belonging to a third party. Rights paperwork is often the weak point: a Data Provenance Initiative review of more than 1,800 text datasets found license information absent for about 70 percent of those on popular hosting sites and wrong for more than half.
The two-verdict record: six fields every opportunity carries
Give every opportunity two verdicts, each with three states, plus the evidence that set it. The data verdict is passed, failed or not tested. The terms verdict is agreed, failed or not reached. Six fields are enough to keep them apart.
A data verdict needs an artifact. Anthropic's guide to agent evals treats a task as good only if two domain experts, working independently, would agree on whether it passed, and NIST's handbook describes acceptance sampling as a decision to accept or reject a lot, not an estimate of its quality. Both point to the same rule: without a test identifier, version and result against a threshold set beforehand, the data verdict is not tested, whatever an email says.
A terms verdict needs a named clause. The Montreal Data License proposal (2019) shows how separable data rights are, distinguishing uses such as research, publishing, internal use and commercializing model outputs. A terms failure coded as scope of use, with the clause and both sides' last positions, can be acted on; one coded as commercial reasons cannot. The six fields of the two-verdict record follow:
| Field | Allowed values | Evidence required | Recorded by |
|---|---|---|---|
| Data verdict | Passed, failed, not tested | Test identifier and version, sampling rule, result against the preset threshold | Acceptance owner |
| Data failure code | Quality defect, coverage or size, task mismatch, documentation | The failing test cases or fields | Acceptance owner |
| Terms verdict | Agreed, failed, not reached | Last written position of each side | Commercial owner |
| Terms failure code | Price, scope of use, duration or exclusivity, risk allocation, authority | The clause in dispute and its draft language | Commercial owner |
| Exit order and dates | Data first, terms first, both, still open | Dated events from request to exit | Deal owner |
| Swap answer | Would pass with other terms, would not, not asked | The counterparty's written reply | Deal owner |
Separate denominators: the same 40 offers, read two ways
Compute each failure rate over the deals that could have produced it: data failures over tested deals, terms failures over deals that reached terms. Mixing denominators is where a funnel analysis quietly goes wrong.
Illustrative: 40 offers reach the sample stage. Nine end on terms before any test, so their data verdict is not tested. Of the 31 tested, 8 fail the acceptance test and 23 pass. Of the 23 that pass, 6 then fail on terms and 17 close. The 8 data failures never reach terms. A single status reports 23 of 40 rejected, or 57.5 percent.
Split by verdict, the picture changes. Data failed in 8 of 31 tested offers, about 26 percent. Terms failed in 15 of the 32 offers that reached them, about 47 percent. The bigger leak is commercial, which a single status would have hidden. The two classic miscounts both distort it: filing the 9 untested deals as data failures gives 17 of 40, or 42.5 percent, overstating data problems by about 17 points, while dropping them from every denominator shrinks the terms failure rate to 6 of 23. The table below sets the two readings side by side:
| Measure (Illustrative, 40 offers) | One rejected status | Two verdicts | What it tells you |
|---|---|---|---|
| Offers lost | 23 of 40 rejected (57.5 percent) | 23 lost: 8 on data, 15 on terms | The headline hides which problem dominates |
| Data failure rate | Not computable | 8 of 31 tested (about 26 percent) | Untested deals stay out of the denominator |
| Terms failure rate | Not computable | 15 of 32 that reached terms (about 47 percent) | Terms are the larger leak in this example |
| Untested offers | Invisible, filed as rejected | 9 of 40 (22.5 percent) | Data quality was never observed for these |
| Common miscount | 17 of 40 data failures (42.5 percent) | Avoided by the not-tested state | Overstates data problems by about 17 points |
Deals that end early are competing outcomes, not missing data
When terms end a deal before testing, the data outcome is not missing at random; it was pre-empted, and the analysis should treat data and terms as competing exits. Treating terms exits as censored, as if the deal might still fail on data later, inflates the estimated chance of a data failure.
Survival analysis has a standard tool for this. The lifelines library's Aalen-Johansen estimator computes the cumulative incidence of each exit type when exits compete, and its documentation warns that treating competing risks as censoring can over-estimate the result: a Kaplan-Meier estimate with competing exits censored describes a world where those exits had been prevented. In deal terms, that is a world where no negotiation ever died on price.
The practical version needs only dated events. For each offer, record the date of the request, the date of the test result if any, the date terms failed or were agreed, and whether it is still open. Then report, at a fixed horizon such as 90 days from request, the share of offers that exited on data, the share that exited on terms, the share that closed and the share still open. Open deals are censored; deals that ended on the other verdict are not.
Is the stated reason the real one? Three cross-checks
A stated reason is a claim, so test it against the record before coding it. Three checks catch most disguised rejections, the price objection that covers a disappointing sample and the quality complaint that ends an awkward negotiation.
The evidence check asks whether a test artifact exists: a data rejection without a result against the agreed threshold is recoded as not tested and routed to terms or to an unknown reason. The order check asks which event came first: if the price objection is dated before the test result, the data cannot have caused the exit. The swap question asks the counterparty in writing whether the records would pass review on different terms, and whether these terms would work if the records passed a named test; its answer goes into the swap field, including not asked.
When the same dataset is offered to several buyers, compare where each exited. Terms failures repeating across buyers point to price or scope of use. A data failure at one buyer while others pass the same records points to that buyer's task rather than the records. Interviewing buyers is a separate method with its own pitfalls; SourceX's guide to diagnosing a buyer's rejection covers the interview itself.
Which fixes reach which failures, including at renewal
Technical work fixes only data failures. Re-preparing records cannot cure a price gap, a scope-of-use disagreement, a supplier without authority to grant the use, or a task the records do not represent, so each code needs an owner who can actually change it.
Route data codes to the delivery team, which can cure a defect, re-sample, narrow coverage or stop. Route price, duration, exclusivity and risk-allocation codes to the commercial owner, who can reprice, shorten the term, drop exclusivity or narrow the grant, for instance to evaluation-only use. Route authority codes to rights review, because missing permission is a fact about the supplier, not a negotiating stance. Route task mismatch back to the requirement, where the honest options are a different source or no deal.
The terms verdict also returns at renewal even when nothing about the data changed. Reddit's annual report for 2024 says that, to date, substantially all of the contract value associated with its licensing revenue is derived from two of its partners, and it warns that those arrangements may not be renewed or may be renewed on less favorable terms. For a data supplier, a renewal is a new terms verdict on data that already passed, and it belongs in the same log.
Routing a lost deal: the two-verdict decision tree
Walk every lost or stalled deal through the same six questions so the record, not the loudest explanation, decides which owner acts next. The tree reads the two verdicts in the order they can be observed.
Run it at exit and again at any reopening, so a deal that failed on terms and later passes a test keeps both histories. The six questions are:
- 1. Was the agreed acceptance test run on this sample or delivery? If not, set the data verdict to not tested and skip to question 4.
- 2. Did the result meet the threshold fixed before testing? If not, set the data verdict to failed, code the failing cases and hand the deal to the delivery owner.
- 3. If it did, set the data verdict to passed and check that the passing result is filed with its test version.
- 4. Were written terms exchanged? If not, set the terms verdict to not reached; the deal is open, or it ended on data.
- 5. Did both sides sign? If not, set the terms verdict to failed and code the clause; if the clause is a use the supplier cannot grant, route it to rights review rather than pricing.
- 6. Ask the swap question in writing and record the answer, including not asked, before closing the file.
Putting both verdicts into the data request
The cleanest way to keep the verdicts apart is to state both sides of the deal in the first request. A requirement that pairs the acceptance test with the intended use defines the data verdict and the scope-of-use terms at once, and makes it obvious later which one failed.
SourceX has no shelf of datasets; it sources from U.S. businesses on request, and the buyer's request-data form names scope and volume alongside language, format, timing and licensing needs, so the data side and the terms side arrive together. On the supply side, SourceX's process has the company approve the scope and terms and sign before any data is delivered, which means a supplier's terms are settled before any buyer-side data verdict can exist. Neither step implies that a dataset will pass a buyer's test or that terms will be agreed; each deal is decided on its own evidence.
A useful request adds three lines to those fields: the acceptance test and its threshold, the intended use stated as a license scope, and which commercial terms are fixed and which are open. A buyer that also says whether evaluation-only use would be acceptable gives the supplier a second path when training terms cannot be agreed.
Frequently asked questions
What is the difference between a data rejection and a failed negotiation?
A data rejection means the records failed an acceptance test the parties agreed in advance. A failed negotiation means they could not agree on price, scope of use, duration, exclusivity or risk allocation. Sales law draws the same line: rejection under UCC 2-601 applies to a nonconforming delivery under an existing contract, while a failed negotiation never produces one.
How should a dataset rejection rate be calculated?
Divide data failures by the number of deals actually tested, not by all opportunities. Deals that ended on terms before testing belong in a not-tested state, outside that denominator. Report the terms failure rate separately, over deals that reached written terms.
Should a deal that ended on price count as a rejected dataset?
No. If price ended the deal before the acceptance test ran, the data verdict is not tested. If the records passed and price ended it afterwards, the data verdict is passed and the terms verdict is failed. Counting either case as a data rejection overstates quality problems.
Can better data preparation fix a terms rejection?
Rarely. Preparation fixes defects, coverage and documentation, which are data failures. Price, scope of use, duration, exclusivity and risk allocation need a commercial change, and a supplier that lacks authority to grant a use needs rights work rather than either.
Which commercial terms decide whether an AI data deal closes?
No public dataset ranks them, so log your own. Code each terms failure as price, scope of use, duration or exclusivity, risk allocation, or authority, with the clause in dispute. Over a few quarters, the codes show which term to change first.
Sources
- UCC section 2-601 lets a buyer reject the whole, accept the whole or accept some commercial units when goods or the tender of delivery fail in any respect to conform to the contract. Source
- NIST's Engineering Statistics Handbook describes acceptance sampling as a decision to accept or reject a lot, whose main purpose is to decide whether the lot is likely acceptable rather than to estimate its quality. Source
- Anthropic's guide to agent evals says a good task is one where two domain experts would independently reach the same pass or fail verdict. Source
- The Montreal Data License proposal (2019) sets out a taxonomy of data rights that separates uses such as research, publishing, internal use and commercializing model outputs. Source
- OpenAI's data partnerships page says it is not seeking datasets with sensitive or personal information or information that belongs to a third party. Source
- The Data Provenance Initiative audited more than 1,800 text datasets and found license information frequently missing (about 70 percent) and wrong (more than half) on popular dataset hosting sites. Source
- The lifelines Aalen-Johansen estimator computes cumulative incidence under competing risks, and its documentation warns that treating competing risks as censoring can over-estimate cumulative incidence. Source
- Reddit's FY2024 Form 10-K says substantially all of the contract value associated with its licensing revenue is derived from two of its partners and that the arrangements may not be renewed or may be renewed on less favorable terms. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.