Skip to content

Deal economics

Dataset pricing by unique decision coverage, not row count

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Price a business dataset on the unique decisions it covers, not its row count. Count distinct decision situations after deduplication, the share linked to recorded outcomes, and the rare exceptions, then frame value as a conditional range that depends on the buyer's task, permitted uses, exclusivity scope, preparation cost and timing. No formula converts coverage into a price.

Value drivers

Dataset pricing unique

This page covers dataset pricing by unique decision coverage, not row count.Illustrative

Key takeaways

  • Start a dataset pricing conversation from unique decisions covered, not from row count.
  • A unique decision records the situation, the choice, the actor or policy, and the outcome, linked by keys.
  • Report raw and deduplicated counts with the method; overcounting is the fastest way to lose a buyer's trust.
  • Rare, outcome-linked exceptions often matter most, but they also concentrate rights and privacy work.
  • Task fit and permitted uses usually move value more than any count.
  • No formula turns coverage into a price; treat value as a conditional range.

How should decision coverage influence business dataset valuation?

Decision coverage should be the starting point of a dataset pricing conversation: count how many distinct decision situations your records capture, with the choice made and the outcome that followed, and how many of those are rare or exceptional. Row count matters only after that, because thousands of rows that repeat the same routine decision teach a model little that a few hundred would not.

There is no formula that converts decision coverage into a price, and no one should promise one. The OECD's work on measuring the economic value of data describes the conceptual and practical challenges of measurement, and research on data valuation in machine learning, such as Data Shapley, defines the value of a training example relative to a specific model and performance measure. Value depends on what a buyer is trying to teach, so a defensible discussion frames it as a conditional range tied to use, rights, costs and timing.

What a supplier can control is the evidence. A decision coverage inventory turns a vague claim that the records are valuable into specific statements a buyer can test: these decision types, this many unique situations after deduplication, this share linked to a recorded result, these exceptions, these permitted uses. That is the asset this page provides, together with a discussion guide for the pricing conversation itself.

What is a unique decision, and how is it different from a row?

A unique decision is one occurrence of a business situation in which someone chose between options under a policy, recorded in a way that shows the inputs, the choice, who or what made it, and what happened next. A row is a storage unit; a decision can span many rows, and many rows can repeat one decision pattern.

At a glance

What is a unique decision, and how is it different from a row

Take a distributor's pricing overrides. In an ERP such as NetSuite, one override can produce a sales order line, an approval workflow entry, an email in the CRM and a later credit memo. That is four or more rows and one decision. Meanwhile, the same ERP may hold a large volume of order lines priced automatically from a list, each a row, none involving judgment. Counting rows treats both the same way; counting decisions does not.

The distinction matters to buyers because models that support or automate work, including agents, need examples of the judgment calls: when to approve, escalate, refuse or ask for more information. Routine rows still have uses, for example as forecasting history, but they answer a different buyer question and should be inventoried separately.

  • Situation: the inputs available at the time, such as customer, item, margin, credit status and request text.
  • Choice: what was decided, from a known set of options.
  • Actor and policy: who decided, under which rule or authority level, and any stated rationale.
  • Outcome: what happened afterward that the system recorded, such as payment, return, callback, churn or dispute.
  • Linkage: the keys that connect the four elements above across systems.

What should the decision coverage inventory include?

The inventory is one row per decision type, filled in from metadata and counts rather than from sharing records. It should be complete enough that a buyer could design a pilot from it and a finance reviewer could see which assumptions drive the range.

Use the template below as is. Fill one copy per decision type, and keep a short note of how each number was produced so it can be re-run later on the delivered data.

The idea is not new. Datasheets for Datasets asks dataset creators to document motivation, composition, collection process and recommended uses so that consumers can judge whether a dataset fits their purpose. A decision coverage inventory applies the same discipline to the one question that matters most in a pricing conversation: which judgment calls the records actually contain.

What should the decision coverage inventory include?
FieldWhat to recordHow to produce itWhy a buyer cares
Decision typePlain-language name and the systems involvedProcess owner interview plus system listMatches the dataset to a model task
Unique situationsCount after collapsing duplicates and templated approvalsGroup by a situation signature, then countSeparates real variety from repetition
Outcome linkageShare of decisions joined to a recorded resultJoin on order, job, ticket or invoice keysOutcomes make decisions checkable
Exceptions and rare casesCount of escalations, overrides, disputes and edge casesFilter on approval level, reason codes, reversalsLong-tail cases are hard to find elsewhere
Rationale captureWhether a reason, note or approver comment existsField fill rate on reason fieldsExplains why, not only what
Policy erasDates when rules, thresholds or systems changedChange logs and process owner inputMixed policies can confuse a model
History spanFirst and last decision date with gaps notedMin and max dates per typeShows stability and seasonality
Permitted usesTraining, evaluation, retrieval or none, per sourceContract and privacy reviewUnlicensable decisions have no value to a buyer
Preparation costEffort to export, link and de-identifyIT and data owner estimateNet value depends on cost and timing

How do you count unique decisions without overcounting?

Count unique situations, not records: collapse rows that belong to one decision, then collapse decisions that are near-identical, and report both the raw and the deduplicated count. Overcounting is the most common way inventories lose credibility once a buyer inspects a sample.

Duplication is not a theoretical concern. Lee and colleagues found that widely used language-model training datasets contain many near-duplicate examples and that removing them reduced memorized output and train-test overlap. Business systems produce their own duplicates: approval workflows that fire twice, emails copied into tickets, recurring orders, and records migrated from an old system and never retired.

  • Collapse all rows sharing one decision key (order line, approval ID, claim number) into one decision.
  • Remove system-generated decisions with no human or policy judgment, such as automatic list pricing, unless the buyer asked for them.
  • Define a situation signature (for example decision type, customer segment, item class, amount band, reason code) and count distinct signatures as a second, stricter measure.
  • Flag near-duplicate free text, such as templated approval comments, and report the share.
  • Separate migrated or backfilled records and state whether they overlap with live-system records.
  • Keep the counting script; the buyer will want the same numbers reproduced on delivered data.

Why do rare decisions carry more weight, and when does row count win?

Rare and exception decisions often carry more weight because they are the cases a model is least likely to have seen and the cases where errors cost the most. A business with a well-documented archive of escalations, disputes and overrides holds examples that are hard to collect any other way.

Machine learning research supports the general point without putting a number on it. Feldman argues that a key difficulty in learning accurate models is having too little data to predict well on rare and atypical instances, which is why memorizing some long-tail examples can be necessary for accuracy. Sorscher and colleagues show that with a good metric for which examples to keep, pruning redundant data can beat the usual power-law relationship between dataset size and error, which implies that not all examples are equally informative.

Two cautions keep this honest. First, rarity only helps if there are enough examples of the rare case for a buyer to learn or test something; a handful of escalations is an anecdote. Second, exceptions are often where personal data, disputes and confidential terms concentrate, so they are also where rights and de-identification review take longest.

Row count matters more when the buyer's task is about patterns in volume rather than judgment, such as demand forecasting, anomaly detection on transactions, or training a model to read a document type at scale. In those cases, long, consistent history and many records per series are the point, and the decision lens adds less.

The assumption with the largest effect on supplier proceeds or buyer fit is usually not the count at all, but whether the buyer's task actually uses the decision type and whether the rights allow that use. A rich decision archive with unclear customer contract terms can be worth less in practice than a plainer dataset that is cleanly licensable. Check rights and task fit before investing in a detailed inventory.

Pricing discussion guide: questions that move the range

Use these questions to run the pricing conversation from evidence rather than from row counts or headline figures. Each one names what strengthens the case, what weakens it and what to bring. None of them produces a price on its own; together they show which assumptions the range depends on.

Pricing discussion guide: questions that move the range
QuestionStronger positionWeaker positionEvidence to bring
Which model task does each decision type serve?A named task the buyer is working onGeneral interest in the dataInventory row mapped to the task
How many unique situations survive deduplication?Large deduplicated count with variety across signaturesLarge raw count that collapses sharplyRaw and deduplicated counts with method
How many decisions link to an outcome?Most decisions joined to a recorded resultOutcomes missing or in a separate unjoined systemLinkage rate per decision type
How well are rare cases covered?Enough exceptions to test or train onExceptions too few or unlabeledException counts by reason code
What uses are permitted?Training and evaluation both clearedRights unknown or restricted by customer contractsRights review notes
Would a pilot confirm utility?Buyer willing to test a scoped sampleNo test plan on either sideProposed pilot scope and success measure
What exclusivity is in scope?Scope defined by decision type, field and termUndefined or blanket exclusivityDraft scope language
What does preparation cost, and when?Export and linkage effort knownUnknown system access or heavy cleanupIT estimate and schedule

Worked example (illustrative): a distributor inventories three decision types

This example is fictional and uses qualitative bands instead of figures, because the point is the method, not a valuation. A mid-size industrial distributor runs NetSuite for orders and invoicing, Salesforce for accounts and quotes, and Zendesk for returns.

Pricing overrides: many raw rows, but after collapsing approval workflow duplicates and automatic list-price lines, the deduplicated count is moderate. Most overrides link to a later invoice and many to a credit memo, and approvers usually leave a short reason. Coverage of unusual situations is good because overrides are exceptions by definition. This type tops the inventory.

Credit holds: fewer decisions, strong outcome linkage because payment history is in the ERP, but rationale is rarely written down. The inventory flags it as useful for outcome prediction and weak for explaining judgment.

Return authorizations: high volume, but most are routine approvals under a published policy and collapse heavily under the situation signature. The minority of disputed returns are valuable, but the free text includes customer names and contract terms, so rights and de-identification review are the deciding factors.

The resulting conversation centers on pricing overrides and disputed returns, proposes a scoped pilot on overrides, and treats routine returns as optional volume. That is a different and more defensible position than leading with the total number of order lines.

How can a buyer pilot test whether decision coverage is real?

A buyer pilot tests the inventory's claims on a scoped, de-identified sample under agreed evaluation terms: does the sample contain the decision types, unique situations, outcome links and exceptions the inventory describes, and does it help on the buyer's task. A pilot turns a supplier's description into shared evidence, which is what moves a range from hypothetical to defensible.

At a glance

How can a buyer pilot test whether decision coverage is real

Agree on the pilot's shape before any records move. The sample should be drawn the way the full delivery would be drawn, not hand-picked, and the counting script from the inventory should be re-run on it so both sides see the same numbers. The success measure should be one the buyer already uses, such as performance on an internal evaluation set, rather than a new metric invented for the occasion.

Write down what happens after each possible result. If the sample confirms the inventory, the conversation moves to scope and terms. If it shows fewer unique situations or weaker outcome links than expected, narrow the offer to the decision types that held up. If the buyer's task turns out not to use the decision type at all, that is useful to learn early and cheaply.

  • Scope: decision types, date range and record families included in the sample.
  • Sampling rule: random within strata, documented, reproducible.
  • Evaluation rights: what the buyer may do with the sample and when it must be deleted.
  • Checks: re-run deduplicated counts, linkage rate and exception counts on the sample.
  • Success measure: the buyer's existing metric on its own task, agreed in advance.

When does a coverage inventory say to drop or shrink the offer?

Stop or narrow when the inventory shows that the decisions you hoped to offer cannot be linked to outcomes, cannot be licensed, or collapse to very few unique situations. Get specialist review whenever personal data, customer confidentiality or financial reporting questions arise.

  • Stop: outcome records live in a system you cannot export, and no keys join them to decisions.
  • Stop: customer or supplier contracts restrict use of the records and cannot be cleared.
  • Narrow: one decision type is strong and the rest are routine; offer the strong one.
  • Narrow: rare cases are concentrated in records with heavy personal data; consider excluding them or scoping de-identification first.
  • Specialist review: counsel for rights, privacy and contract questions; a finance reviewer for how any proceeds, costs and timing are treated; the commercial lead for exclusivity scope.
  • Re-run: refresh the inventory when buyer requirements, rights or costs change.

How SourceX uses decision coverage when assessing a dataset

SourceX's process for licensing business records as AI datasets runs from sourcing and qualification through rights review, contracting, data preparation, delivery and payment, and decision coverage matters most at the qualification step. The general guide it publishes is two or more years of consistent history and tens of thousands of records in one system, with room for smaller sets when the work is rare and every dataset judged on its own terms.

That last point is where decision coverage comes in: a smaller archive of rare, outcome-linked decisions can be worth a conversation even when the row count looks modest. Before any of that, SourceX looks at the company itself through a short assessment, and no records are uploaded at that stage.

Frequently asked questions

What is decision coverage in a dataset?

Decision coverage is the range of distinct business situations a dataset records, each with the choice made and, ideally, the outcome that followed. It is measured after duplicates and automatic decisions are removed. High coverage means a model can see many different judgment calls, not the same one repeated.

Does more data always mean a higher dataset price?

No. Duplicated or routine rows add little for many AI tasks, and research on data pruning shows that removing redundant examples can preserve or improve results. Volume matters most for tasks like forecasting that depend on long, consistent series.

Which assumption has the largest effect on what a dataset is worth to a buyer?

Usually whether the buyer's task actually uses the decision type and whether the rights allow that use. A rich archive that cannot be licensed, or that does not match a buyer's work, is worth little to that buyer regardless of size.

How do I count unique decisions without sharing records?

Run counts inside your own systems: collapse rows sharing a decision key, drop automatic decisions, and count distinct situation signatures. Share the counts and the method, not the records. Keep the script so the numbers can be reproduced on delivered data.

Should exceptions and disputes be included in a licensed dataset?

They are often the most useful records, but they also concentrate personal data, confidential terms and disputes. Include them only after rights review and a de-identification plan, and be prepared to exclude categories that cannot be cleared.

Sources

  • The OECD's note Measuring the economic value of data clarifies what data means for valuation purposes and identifies the conceptual and practical challenges of measuring its value. Source
  • Ghorbani and Zou propose Data Shapley as a metric for the value of each training datum to a predictor's performance and develop methods to estimate it. Source
  • Lee et al. found common language-model datasets contain many near-duplicate examples and that deduplication reduced memorized output and train-test overlap. Source
  • Feldman argues that a primary hurdle to accurate learning is insufficient data on rare and atypical instances, making memorization of long-tail examples necessary for accuracy. Source
  • Sorscher et al. show that pruning training data with a high-quality ranking metric can make error fall faster than a power law in dataset size. Source
  • Datasheets for Datasets proposes documenting each dataset's motivation, composition, collection process and recommended uses. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies