Skip to content

Deal economics

Stress five risks at once in a dataset licensing downside model

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A dataset licensing downside model should price five losses in the order they can occur: a buyer that never commits, a delivery the buyer rejects, refunds after acceptance, a rights challenge that reaches every buyer of the same records, and exclusivity that removes resale. In this article's illustrative model, each stress alone leaves most of the expected contribution; four together with an acceptance-order mismatch leave almost none.

License terms

Stress five risks

This page covers stress five risks at once in a dataset licensing downside model.Illustrative

Key takeaways

  • Order decides cost: losses that land before money is spent are cheap, while refunds and rights challenges land after payment and can repeat across buyers.
  • Owing the supplier before the buyer accepts cut the rejection rate an exclusive deal could absorb from 72.0% to 47.2% in the illustrative model.
  • A rights challenge can unwind every license sold from the same records; in Thomson Reuters v. Ross, data obtained through an intermediary left the buyer defending the claim.
  • Exclusivity raises the first price but removes the resale cushion: with no downside events, the illustrative exclusive deal earns 60.0 instead of 75.0.
  • Single stresses left expected contribution between 30.8 and 41.3 against a base of 44.5; the joint stress left 0.6.

What should a dataset licensing downside model answer?

It should answer one question: which combination of plausible failures turns a signed licensing deal from profitable to loss-making, and how much evidence the company holds about each failure. A single expected-value line cannot answer that, because the failures share causes. A defect found late drives both rejections and refunds; a flaw in the rights drives refunds for every buyer at once.

For an investor the useful output is a map, not a forecast. It shows which losses land before money is spent and which land after it, how much of a deal's contribution survives each one, and where the break-even line sits when several arrive together. A company whose terms surface failures before money is spent has a different risk profile from one whose failures surface after payment, even at identical prices.

This is a deal-level model. It does not value a company or estimate whether a business could be permanently impaired; it tests whether one licensing deal, or a book of similar deals, keeps a positive contribution when acceptance, refunds, rights and exclusivity go wrong at the same time.

Five losses in the order they can happen: the downside register

Five losses matter, and their order decides how much each costs. A buyer that never commits wastes only the work done before commitment. A rejected delivery wastes preparation as well. A refund after acceptance hands back cash the company may already have passed to its supplier. A rights challenge can unwind every sale of the same records, and exclusivity, agreed at signing, removes the resale income that would otherwise absorb the other four.

Read the register by its second and third columns. Losses that can only occur before money is spent are cheap to absorb; losses that occur after payment, or that repeat across buyers, are the ones a committee should price carefully:

Five losses in the order they can happen: the downside register
EventWhen it can occurAlready spent when it hitsWhat is lostEvidence to request
No closingBefore the buyer signs or commitsRights review, plus any preparation started earlySunk review and preparation; no revenueProposed versus signed deal counts, and which costs start before commitment
Acceptance failureAfter delivery, during the buyer's reviewRights review and preparation; the supplier's fee if it was already owedAll of that, unless the dataset is fixed or sold to someone elseAcceptance criteria in each buyer contract and a rejection log with reasons
Refund or reversalAfter acceptance and invoicingEverything, including any supplier paymentThe refunded share; a supplier fee is rarely recoverableRefund and credit clauses and every refund issued, by cause
Rights challengeAny time in the license life, often after trainingEverything, sometimes including the buyer's model workRefunds across every buyer of the records, defense cost, indemnity claimsChain-of-title file, warranties and indemnity caps in both directions
Exclusivity limiting resaleAt signing, when a buyer asks for itNothing yetThe resale income that would absorb the other four lossesExclusivity scope, field of use, term and the premium paid for it

Why the order of acceptance can matter more than the rejection rate

A rejected delivery costs a few points when the seller owes its own supplier nothing yet, and much more when it does. In the illustrative model, owing the supplier before the buyer accepts lowers expected contribution only from 44.5 to 41.3 at a 10% rejection rate, but it moves the break-even point sharply: for an exclusive sale with no refunds, the rejection rate that wipes out expected contribution falls from 72.0% to 47.2%.

The reason is mechanical. When the two acceptances are aligned, a rejection costs preparation and review; when they diverge, it costs the data as well, and the data is usually the largest cost in the deal. A higher rejection rate with matched terms can be safer than a lower one with mismatched terms.

So the first diligence question is about sequence, not rates: does the company become liable to its supplier before, at the same time as, or after its buyer accepts? Back-to-back terms, conditional supplier payment and lot-by-lot acceptance each shorten the mismatch, and each should appear in the documents, not only in the pitch.

Rights challenges reach every buyer of the same records

A rejection hits one sale; a successful rights challenge hits every license sold from the same records, adds defense cost and can reach the buyer's models. That is why the rights row in the register is the one most likely to be correlated across a whole book of deals.

Thomson Reuters v. Ross Intelligence shows the chain. The District of Delaware's February 11, 2025 opinion records that "to train its AI, Ross made a deal with LegalEase to get training data in the form of 'Bulk Memos.'" It found that Ross infringed 2,243 Westlaw headnotes. On interlocutory appeal, the Third Circuit's opinion in No. 25-2153 held the headnotes original and found the first fair-use factor against Ross, writing that its use "shares the same ultimate purpose as Thomson Reuters's use, making ROSS's use minimally transformative, at best." The data came through an intermediary; the challenge landed on the buyer.

Outcomes are hard to forecast in both directions. In Getty Images v Stability AI ([2025] EWHC 2863 (Ch), 4 November 2025), the training and development claim was abandoned because, as the judgment records, there was no evidence that training and development of Stable Diffusion took place in the United Kingdom. Where and how training happened, not only what the data contained, can decide a challenge, which argues for modeling rights risk as a range rather than a point.

Severity can exceed a refund. The FTC's 2022 stipulated order against WW International and Kurbo defined "Affected Work Product" as "any models or algorithms developed in whole or in part using Personal Information Collected from Children through the Kurbo Program" and required it destroyed. On the contract side, Getty Images described its September 2023 generative AI license as carrying "uncapped indemnification", a reminder that the indemnity cap a licensor accepts sets how much of this tail it keeps.

Exclusivity buys a higher price and gives up the cushion

Exclusivity trades a higher first price for the resale income that would otherwise absorb rejections, refunds and challenges. In the illustrative model, an exclusive sale at a 20% premium with no resale lowers expected contribution from 44.5 to 33.4, and with no downside events at all the exclusive deal earns 60.0 instead of 75.0.

Two kinds of exclusivity point in opposite directions. Supply-side exclusivity, where the source company promises not to license the same records to others for the same use, protects a seller's resale value. Buy-side exclusivity, where the onward buyer demands sole use, removes it. A memo should record which kind each deal carries, its field of use and term, and the premium actually paid, rather than treating exclusivity as a single yes or no.

Resale also interacts with rights. Each extra buyer adds income, but each one would need a refund if the records' rights failed, so in the model the rights exposure per dataset grows with every resale. A seller with many non-exclusive buyers of one dataset holds more cushion and a larger correlated tail at the same time.

Stressed alone, then together: the illustrative downside model

Each single stress leaves most of the expected contribution: moved one at a time, the six stresses lower it from 44.5 to between 30.8 and 41.3. Applied together, a 30% rejection rate, supplier liability before buyer acceptance, refunds of 15% of contract value, a 10% rights-challenge rate and an exclusive sale leave 0.6. Their single-stress changes sum to 43.0 points; together they cost 43.9, because the stresses compound.

Inputs, all illustrative and in percent of the first buyer's contract value (100): a data cost of 40 owed once the seller closes with its supplier, preparation of 15 spent before acceptance, rights review of 5 spent before closing, one resale at 40 with a marginal cost of 5, and a defense cost of 15 if rights are challenged. Base rates: 20% of deals never close, 10% of deliveries are rejected, refunds average 5% of contract value, and 3% of datasets face a successful rights challenge over the license life. None of these are SourceX's terms, prices or rates.

Expected contribution weights the accepted outcome (net of refunds, resale income and expected rights cost) and the rejected outcome by the rejection rate, scales the result by the share of deals that close, and subtracts the review cost of deals that never close. The break-even frontier in the figure plots, for each rejection rate, the refund share at which expected contribution reaches zero: at a 30% rejection rate it is 60.0% for a non-exclusive deal with one resale, 38.0% for an exclusive deal and 23.7% for an exclusive deal whose supplier is owed first.

The table lists each scenario so a reader can rerun it with a company's own evidence:

Stressed alone, then together: the illustrative downside model
Scenario (illustrative)Input changedExpected contributionChange from base
Base caseNone44.50.0
No closing20% to 40% of deals32.1-12.4
Acceptance failure10% to 30% of deliveries30.8-13.7
Acceptance orderSupplier owed before the buyer accepts41.3-3.2
Refunds or reversals5% to 15% of contract value37.3-7.2
Rights challenge3% to 10% of datasets36.6-7.8
ExclusivityNo resale, 20% price premium33.4-11.1
Joint stressRows three to seven together0.6-43.9
A line chart of the break-even refund share against the acceptance-failure rate from 0 to 60 percent for three deal structures; at a 30 percent failure rate the break-even refund share is 60.0 percent for a non-exclusive deal with one resale, 38.0 percent for an exclusive deal and 23.7 percent for an exclusive deal whose supplier is owed before the buyer accepts.
Refund share at which expected contribution reaches zero. Illustrative model inputs, not any company's terms or rates.

Is this a probability tree with extra steps?

No. A probability tree multiplies branch probabilities and reports one expected value; this model asks which costs are already spent when each loss lands and stresses losses together, because one cause usually moves several at once. The probabilities are inputs to be argued over, not the output.

The strongest alternative explanation is that the downside is small because the seller passes every loss to its suppliers. Test it in the documents rather than accept it. Is supplier payment conditional on the buyer's acceptance? Can a refund be recovered from the supplier, and does the supplier have the money? Do the warranties the seller gives buyers match the warranties it receives? If any answer is no, the pass-through holds only for some losses.

Forecasts deserve the same suspicion. A revenue forecast that counts signed but unaccepted deals at full value, and treats rejections, refunds and rights claims as independent small percentages, will look safest exactly when the losses are most correlated. Ask for the forecast rerun under the joint stress.

What would make this downside model wrong

The model is a frame for evidence, and these are the ways it can mislead:

  • No public base rates. Rejection, refund and rights-challenge rates for AI data deals are not published, so every rate here is a placeholder until a company's own logs replace it.
  • Legal outcomes cut both ways. Ross and Getty point in different directions on different facts and laws; the rights rate could be far lower or far higher than modeled.
  • Indemnities and insurance move the tail. A capped indemnity, a supplier warranty with a solvent supplier or an insurance policy can shrink the rights row; an uncapped indemnity can enlarge it.
  • Prices can fall. More sellers, synthetic data or buyers sourcing in-house can lower contract values and resale prices, shrinking every cushion at once.
  • Buyers can move together. When few buyers dominate, one buyer's rejection can coincide with others pulling back, which the per-deal model treats as independent.
  • Timing is ignored. The model counts contribution, not cash; a deal that breaks even on paper can still strain cash if refunds arrive before resale income.
  • Privacy enforcement has a long tail. Orders like Kurbo's reach models built on the data, a severity no refund line captures.

Reading SourceX's published terms through the model

SourceX's published terms place several of these losses, which makes it a useful worked application. No records move until the supplier has signed off on scope and terms, and from then on the signed agreement governs how SourceX may use, sublicense and distribute the de-identified dataset. Its how-it-works page says it seeks commitment on approved terms and signs its own onward agreement before instructing processing, which puts preparation after commitment, and that data which never reaches Closing earns the supplier nothing.

On exclusivity, the published terms describe the supply-side kind: a supplier commits not to license the selected records to anyone else for AI training, with the term and any alternatives set in that deal's Dataset Deal Confirmation, which is not published, and the commitment restricts the supplier rather than SourceX's ability to distribute or sublicense the de-identified dataset. Onward buyer terms, acceptance criteria, refund history and resale counts are not public; SourceX shares company and commercial metrics with investors directly.

Investors assessing SourceX's raise, open to accredited investors under Rule 506(c), can request materials at sourcex.si/investors and test these questions against them. They fit any data licensing company:

  • Can a supplier's Closing happen before the onward buyer accepts, and if so, how often has it?
  • What share of deliveries were rejected or partly accepted, by cause and by cohort?
  • What refund, credit or reversal obligations do onward licenses carry, and which have been triggered?
  • How many buyers hold licenses to the same records, and what indemnity cap applies per dataset?
  • How often do onward buyers ask for exclusivity, and what premium and field limits come with it?
  • Which costs, if any, start before an onward commitment is signed?

Methods note: inputs, formula and what was not measured

The downside register, the scenario table and the break-even frontier are editorial frameworks built for this article on 2026-10-27. All inputs are illustrative and expressed in percent of the first buyer's contract value; they describe no real company, deal, price or rate, and none is a SourceX term or probability.

Formula: expected contribution = (1 - no-closing rate) x [(1 - rejection rate) x accepted outcome + rejection rate x rejected outcome] - no-closing rate x review cost. The accepted outcome is price x (1 - refund share) - data cost - preparation - review + resales x (resale price - resale cost) - rights rate x (price + resales x resale price + defense cost). The rejected outcome is the loss of preparation and review, and of the data cost as well when the supplier was owed first. Arithmetic used exact fractions, each result rounded once, half up, to one decimal; the frontier solves the linear equation for the refund share at rejection rates from 0% to 60% in steps of 5, floored at zero.

Public evidence comes from the District of Delaware and Third Circuit opinions in Thomson Reuters v. Ross Intelligence, the High Court judgment in Getty Images v Stability AI, the FTC's stipulated order in the WW International and Kurbo matter, and Getty Images' September 2023 product announcement, each checked through domain-limited search excerpts on 2026-10-27. Not measured: real rejection, refund, resale or rights-challenge rates at SourceX or any data company, and the final disposition of the Ross appeal, which the verified excerpt did not show.

Frequently asked questions

What is a downside model for a dataset licensing deal?

It is a deal-level test of how much contribution survives when the deal goes wrong: the buyer never commits, rejects the delivery, claws back money, faces a rights challenge or demands exclusivity. It records when each loss can occur relative to money already spent, then stresses the losses alone and together to find the break-even line.

Why does the order of acceptance matter in a data licensing deal?

Because it decides whether a rejection costs preparation alone or the data as well. If a seller becomes liable to its supplier before its buyer accepts, a rejected delivery leaves it paying for data it cannot sell to that buyer. In this article's illustrative model, that mismatch cuts the rejection rate an exclusive deal can absorb from 72.0% to 47.2%.

How does a rights challenge affect a dataset sold to several buyers?

It can hit all of them at once. A rejection affects one sale, but a successful challenge to the records' rights can require refunds or indemnity payments to every licensee, plus defense costs, and in privacy cases regulators have ordered models built on the data destroyed. That correlation makes the rights row the largest tail in most books of deals.

Does exclusivity make a data licensing deal safer?

Not automatically. A buyer's exclusivity usually raises the first price but removes resale income that would absorb other losses. A supplier's exclusivity toward the seller does the opposite, protecting resale value. Record which kind a deal carries, its scope and term, and the premium actually paid.

Where should the inputs for a downside model come from?

From the company's own documents: rejection logs with reasons, refund and credit records by cause, buyer and supplier contracts showing acceptance and payment sequence, chain-of-title files and indemnity caps. Until those exist, every rate in the model is a placeholder, and the model's job is to show which evidence matters most.

Sources

  • The District of Delaware's February 2025 opinion in Thomson Reuters v. Ross Intelligence records that Ross obtained training data from LegalEase as Bulk Memos and found Ross infringed 2,243 headnotes. Source
  • The Third Circuit's precedential opinion in No. 25-2153 held Westlaw headnotes original and found the first fair-use factor weighed against ROSS's minimally transformative use. Source
  • In Getty Images v Stability AI ([2025] EWHC 2863 (Ch)), the training and development claim was abandoned for lack of evidence that training took place in the United Kingdom. Source
  • The FTC's stipulated order against WW International and Kurbo defined affected work product as models or algorithms developed using personal information collected from children and required it destroyed. Source
  • Getty Images announced in September 2023 that its generative AI offering's license includes uncapped indemnification. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies