Skip to content

Deal economics

Dataset valuation transferability: will it work outside your systems?

By SourceX Editorial · Updated

9 sections

Short answer

Export a sample exactly as an outsider would receive it, then hand it with its documentation to someone who has never used your systems. If they can decode every code, rejoin linked records, tell staff entries from system defaults and reproduce each recorded outcome, the data is portable. Then ask whether its patterns would hold at other companies. Value that needs your software or your staff to explain it does not travel.

Value drivers

Dataset valuation

This page covers dataset valuation transferability: will it work outside your systems.Illustrative

Key takeaways

  • Value that depends on your software or staff to explain it does not survive export.
  • Portability (can a stranger use it) and generalizability (does it hold elsewhere) are separate tests.
  • A cold reader with only the export and its documentation is the most direct evidence of portability.
  • State value as a range tied to assumptions; a defined buyer task with recorded outcomes usually matters most.
  • Company-specific data can still fit buyers who want rare work or procedure-following evaluations, if local practices are labeled.

How do you test whether a dataset keeps its value outside your own systems?

Test the export, not the system. Inside your software, a record is surrounded by help: code tables that turn 7 into Warranty callback, screens that show the linked invoice, staff who know that a blank field means the technician forgot rather than nothing happened. An AI developer receives none of that. What they receive is files and whatever documentation travels with them.

So the test has two halves. Portability asks whether a capable stranger can understand and use the records with only the export and its documentation. Generalizability asks whether what the records teach holds beyond your company, or only describes how your team happens to work. A dataset can pass the first and fail the second, and the two failures call for different fixes.

This guide gives you an audit for both, a way to run it with a cold reader, and a framework for discussing value as a range that depends on stated assumptions. It does not estimate prices, and no single test does.

Why does value that is obvious inside your company disappear on export?

Because much of the meaning lives in the system and the people around the data, not in the rows. The OECD's work on measuring the economic value of data separates two components of value: content, the information the data embodies, and context, the circumstances in which the data were gathered and could be used. It also notes that data volume, though easy to count, is unlikely to be a good basis for inferring value. A large export that loses its context can be worth less than a smaller one that keeps it.

A later OECD paper adds that the value of data depends to a large extent on the governance framework that determines how it can be created, shared and used. For a business, that means rights and permitted uses are part of transferability too: records that cannot leave under your customer contracts have no value to an outside developer, however well documented they are.

The usual losses on export are predictable: internal codes with no lookup table, custom fields whose meaning only one administrator remembers, links that depended on the user interface, statuses whose meaning changed when the workflow changed, and free-text shorthand that only your staff can read.

Portable versus generalizable: which kind of transferability are you testing?

Keep the two questions apart, because buyers weigh them differently depending on what they are building. A team building an evaluation of a model's ability to follow one company's procedures cares mainly about portability. A team training a model meant to work across many companies cares about generalizability as well.

Portable versus generalizable: which kind of transferability are you testing?
DimensionQuestion it answersEvidence you can produceWhat a failure changes
ReadabilityCan a stranger decode every field?Data dictionary, code tables, a cold-reader logFixable before any sale with documentation work
LinkageDo related records rejoin without your software?Stable keys and a join test on the sampleFixable if keys exist; serious if links lived only in the interface
Outcome clarityIs the result of each case recorded and unambiguous?Outcome field definitions and a sample auditWithout outcomes, many training and evaluation uses fall away
Stability over timeDid meanings change when systems or forms changed?A change log with datesRequires splitting the data into consistent periods
GeneralizabilityWould the patterns hold at other companies?Comparison with industry practice, notes on what is localNarrows the set of buyers rather than ending the conversation

The transferability and documentation audit

Run each check on a representative sample of the export, not on the live system. The pass condition is written so that someone other than the auditor can confirm it. Version 1.0 of this audit is dated 2026-10-18; adapt the checks to your record types, but keep the pass conditions concrete.

At a glance

The transferability and documentation audit

Two open specifications are useful models for the documentation side. Frictionless Data's Table Schema gives each field a name, type, description, constraints and optional categories, and lets a schema declare which strings count as missing values and which foreign keys link tables. Croissant, the MLCommons metadata format for machine learning datasets, describes record sets and fields and can carry provenance at several levels. Datasheets for Datasets, from Gebru and colleagues, proposes that every dataset come with documentation of its motivation, composition, collection process, recommended uses and more.

The transferability and documentation audit
CheckHow to run itPass conditionIf it fails
Codes resolveList every coded field and look up each distinct valueEvery value appears in an exported lookup table with a plain-language labelExport the lookup tables or map codes to labels
Fields documentedCompare exported columns with the data dictionaryEvery column has a type, description and allowed valuesWrite the missing entries with the system owner
Missing values explainedCount blanks and placeholder values per fieldEach placeholder is declared, and blanks have a stated meaningDeclare missing-value markers; note fields where blank is ambiguous
Links rejoinJoin related tables on exported keysThe joins reproduce the relationships shown in the source system for sampled casesExport keys; flag relationships that existed only on screen
Outcomes recordedPick sampled cases and find the final resultA stranger finds the outcome for each case without asking anyoneAdd the outcome source or narrow to cases that have one
Entry source knownIdentify system defaults and auto-filled fieldsHuman-entered and system-generated values can be told apartDocument defaults and automation rules
Changes datedList form, workflow and system changes over the periodEach change has a date and the fields it affectedSplit the data into consistent periods
Shorthand readableRead a sample of free text with a cold readerAbbreviations are covered by a glossaryBuild a glossary from the most frequent terms
Local practice flaggedAsk which procedures are unique to your companyCompany-specific rules are written down as suchLabel them so a buyer can judge fit
Rights recordedMap each source system to the agreements that govern its recordsEach system has a stated basis for licensing its recordsExclude records whose use cannot be established

Running the cold-reader test

The cleanest evidence of portability is a person outside the team succeeding with only the export. Choose someone technical who has never used your systems: an analyst from another department, an outside data engineer under a confidentiality agreement, or an advisor. Give them de-identified sample files and the documentation, and nothing else.

  • Ask them to answer five to ten concrete questions about sampled cases, such as what happened, how long it took and why it was closed.
  • Log every question they have to ask your team; each one is a documentation gap.
  • Record which answers they get wrong and why: a missing code, a broken link, an ambiguous status.
  • Fix the documentation, then repeat with a different reader until the log of questions is short and specific.
  • Keep the final log; it is evidence a buyer can read, and it shows the work behind the data dictionary.

Which assumption moves value the most?

Value for AI use is conditional, so state it as a range tied to assumptions rather than a single figure. The assumption that usually dominates is whether the records support a buyer's specific task with a recorded outcome; completeness, rights, preparation effort, timing and the buyer's alternatives adjust the range around it. The WILDS benchmark, built by researchers to study distribution shift in real deployments, reports that standard training gives substantially lower performance on held-out hospitals, locations or time periods than on the data it was trained on, which is why a buyer may ask how typical your operation is.

Write the assumptions down with what changes if each one fails. That turns a valuation conversation into a checklist both sides can test.

Which assumption moves value the most?
AssumptionIf it holdsIf it fails
Records support a defined buyer task with outcomesThe data can be scoped to that taskOnly descriptive uses remain, and buyer interest narrows
The export is portablePreparation effort is mostly de-identification and packagingDocumentation work comes first, and timing moves out
Patterns are typical of the industryThe set fits buyers training general modelsIt fits buyers who want this niche or need evaluation data
Rights to license are clearThe dataset can be offeredAffected records are excluded, whatever their quality
Volume and history are sufficientThe set stands on its ownIt may still qualify if the work is rare

Worked example (illustrative): work orders at a field-service company

This example is invented to show the method. A commercial HVAC service company exports three years of work orders from its field-service platform. The audit finds 42 distinct status codes, of which 9 have no exported label; 14 custom fields, of which 4 are undocumented; and a parts table that rejoins to work orders only through a reference number visible on screen but not included in the default export.

A cold reader, given the sample and the first draft of the dictionary, raises 23 questions on the first pass. Most concern the unlabeled statuses and one custom field whose meaning changed after a form redesign in the second year. After the team exports the lookup tables, adds the parts key and dates the form change, a second reader raises 4 questions, all about technician shorthand, which a short glossary resolves.

On generalizability, the team notes that its callback rule, a free return visit within a set number of days, is company policy rather than industry practice, and labels it as such. The export is now portable; whether it generalizes depends on the buyer's task, and the documentation lets the buyer decide.

The exception: when company-specific data is exactly what a buyer wants

Low generalizability is not always a defect. Some buyers want records from a narrow operation precisely because the work is unusual: rare equipment, a specialized regulatory process, or exceptions that larger datasets lack. Others build evaluations that test whether a model can follow one organization's documented procedures, where the local rules are the point.

In those cases the documentation matters even more, because the buyer must be able to tell what is local from what is general. The audit's local-practice check exists for this: it lets a narrow dataset be offered honestly instead of being presented as typical.

Where the audit fits in a SourceX assessment

Owners can start before any of this work is finished: SourceX's first step is a short company assessment with no data uploads, so the audit can come after that first conversation rather than before it. Buyers tend to look for consistent history, often two years or more, and tens of thousands of records in one system, though smaller sets can qualify when the work is rare, and each dataset is assessed on its own. A completed transferability audit gives the rights review, preparation and qualification work SourceX does before records are licensed to AI developers a documented starting point.

Frequently asked questions

What makes a dataset valuable to AI developers outside the company that created it?

Records that a stranger can decode, rejoin and read outcomes from, covering a task the developer cares about, with a clear right to license. Volume helps only when that context survives the export.

Is a data dictionary enough to make an export portable?

It is necessary but rarely enough. A dictionary describes fields; portability also needs exported lookup tables, working keys between tables, dated records of workflow changes and a glossary for free text. The cold-reader test shows what is still missing.

Does company-specific data have less value?

Not necessarily. It may fit fewer buyers training general models, but it can suit buyers who need rare work or evaluations of procedure-following. Label local practices clearly so the buyer can judge fit.

How large a sample should the audit use?

Large enough to include every record type, status and period in the export, including rare cases. A sample drawn only from recent, typical records will miss the codes and changes that cause most failures.

Can the audit produce a price for the dataset?

No. It tells you which assumptions about value hold and which need work. Price depends on a buyer's task, rights, preparation effort and alternatives, and is set deal by deal.

Sources

  • The OECD identifies content and context as components of data value and says data volume is unlikely to be a good basis for inferring value. Source
  • The OECD finds that the value of data depends to a large extent on the governance framework determining how it can be created, shared and used. Source
  • Frictionless Table Schema gives fields a name, type, description, constraints and categories, and lets a schema declare missing values and foreign keys. Source
  • Croissant 1.1 describes datasets through record sets and fields and supports provenance at dataset, record set, field and value level. Source
  • Datasheets for Datasets proposes that every dataset be documented for its motivation, composition, collection process and recommended uses. Source
  • The WILDS benchmark of real-world distribution shifts, across hospitals, locations and time, shows a clear gap between in-distribution and out-of-distribution performance. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies