Skip to content

Deal economics

Customer exclusions in data licensing: does buyer fit survive?

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Test buyer fit by profiling the dataset twice, with and without the largest customer, and comparing what a buyer evaluates: volume, years of history, task type coverage, outcome labels, rare cases and customer spread. Fit survives when every needed property stays above its threshold. It fails when the excluded account held most of one category or the hard cases, even if total volume still looks large.

Value drivers

Data licensing customer

This page covers customer exclusions in data licensing: does buyer fit survive.Illustrative

Key takeaways

  • Record concentration, not revenue concentration, decides how much an exclusion costs in buyer fit.
  • Compare at least three scopes: full, minus the customer field match, and minus every other mention of the account.
  • Set pass thresholds from the use case before computing the numbers, so the table cannot be tuned afterward.
  • An exclusion can improve fit when one account dominates and the categories a buyer needs survive.
  • Proof is an exclusion register with filters, before and after counts and a zero-hit leakage search, rerun on the final export.

How do you test buyer fit after removing your largest customer's records?

Build the dataset profile twice, once with every eligible record and once without the largest customer, and compare the two on the properties an AI buyer actually evaluates: volume, years of history, coverage of task types, outcome labels, rare cases and the spread of records across the remaining customers. Buyer fit survives the exclusion when every property the buyer cares about stays above its threshold. It fails when the excluded account held most of one task type, one region, one product line or most of the hard cases, even if the total record count still looks large.

The output is a scope sensitivity and case coverage analysis: a short table that shows, for each property, the value with and without the account, the change, and whether the change breaks a claim you would make to a buyer. Leadership can then decide whether a scoped investigation is still worth running, using conservative numbers and a written list of what is not yet known.

The test is cheap because it runs on metadata. You need record counts, dates, category fields and customer identifiers from the source system, not the record contents, so it can be done before anyone exports text, attachments or personal data.

Why one account can carry more of the dataset than its share of revenue

Revenue concentration and record concentration are different numbers, and the second is the one that matters here. A customer that is a modest share of revenue can generate a large share of tickets, work orders or transactions if it buys high-frequency, low-value services, uses a dedicated portal, or runs a contract with strict reporting that forces every event to be logged. The reverse also happens: a large account billed through one annual contract may leave few operational records.

Public companies already think about the revenue side. SEC Regulation S-K Item 101 asks registrants to describe any dependence on key customers in their business description. A data licensing decision needs the same discipline applied to records: how dependent is the dataset, not the income statement, on one source?

A simple way to put a number on it is the Herfindahl-Hirschman Index, which the U.S. Justice Department describes as the sum of the squared market shares of each firm in a market, approaching zero when many firms of similar size compete and reaching 10,000 when one firm controls everything. Applied to a dataset, each customer's share of records replaces market share. The antitrust thresholds do not transfer to datasets, so use the index only to compare scopes against each other: if removing one account sharply lowers the index, the remaining data is more evenly spread, which some buyers prefer; if the count of usable records collapses at the same time, the gain in balance may not be worth it.

Separate the reason for the exclusion from its effect. A customer may be excluded because its contract has confidentiality or data-use terms, because it asked, because its records would reveal its identity, or because leadership does not want the relationship touched. Those are questions of contract, privacy and relationship risk, and counsel reviews the contract language. This analysis only measures what the exclusion costs in buyer fit once the decision is made or being considered.

What should the scope sensitivity and case coverage analysis include?

Use one row per property a buyer would test, and compute each value for the full eligible scope and for the scope without the excluded account. The pass rule column turns the numbers into a decision; set the thresholds from the buyer request or use case before you look at the results, so the table cannot be tuned after the fact.

What should the scope sensitivity and case coverage analysis include?
PropertyHow to measure with and without the accountMaterial loss whenEvidence kept
Usable record volumeCount records that pass basic eligibility filters (complete, in date range, not already excluded for other reasons)The remaining count falls below what the target use case needsSaved count query with filter definitions and run date
History depthFirst and last record date, and records per year, per scopeWhole years drop out or one year becomes thin enough to break a trendYear-by-year count table for both scopes
Task type coverageRecords per task category (ticket type, job type, transaction type) and share of the totalA category the buyer named falls below a minimum count or disappearsCategory count table with the source field named
Outcome labelsShare of records with a verified outcome (resolved, paid, completed, reversed), per scopeLabeled share drops because the account had the cleanest closure dataField definitions showing how outcome is recorded
Rare and hard casesCount of escalations, disputes, rework, exceptions or other flagged edge casesMost edge cases came from the excluded accountEdge-case definition and counts
Customer spreadNumber of remaining customers, top-five share of records and a concentration indexOne remaining customer now dominates, or too few customers remain to generalizeCustomer-level counts with identifiers hashed
Segment coverageRecords by region, industry served, product line or channelA segment the buyer asked for is now covered by one or two customersSegment count table
Generalization checkHold out whole customers in evaluation, if a baseline model is runPerformance on held-out customers falls well below in-sample performanceSplit definition and evaluation notes
Identity leakageSearch remaining records for the excluded customer's names, domains, sites and contract numbersMentions of the excluded customer remain in other customers' records or in free textSearch terms list and hit counts

Which comparison scopes reveal the real cost of the exclusion?

Run at least three scopes, not two. The first is the full eligible scope. The second removes every record whose customer field matches the excluded account. The third also removes records that mention the account elsewhere, such as subcontractor notes, shared sites, email threads copied into tickets, or invoices to an affiliate. The gap between the second and third scope is often where surprises sit, because the customer field is rarely the only place an account appears.

If the exclusion is partial, add a fourth scope. Some companies can keep an account's structured fields, such as dates, categories and durations, while dropping its free text and attachments. Measure that scope separately, because it may keep most of the coverage while removing most of the identity risk. Whether partial use is allowed is a contract question, not an analytics one.

When a baseline model is part of the evaluation, group the split by customer. The scikit-learn documentation describes GroupKFold as a variation of k-fold that keeps the same group out of both the training and test sets, and LeaveOneGroupOut as holding out the samples of one group in each split. Holding out whole customers tells you whether patterns learned from the remaining accounts carry over to an unseen one, which is closer to how a buyer will use the data. The WILDS benchmark paper is a useful warning here: across its real-world datasets, models trained in the standard way performed substantially worse on data from shifted conditions than on data like their training set.

Worked case (illustrative): a facilities maintenance company and its university account

This case is invented to show how the table is filled in; every count and percentage is illustrative, not an observation. A regional facilities maintenance company has seven years of work orders in its field service system. Its largest account, a university system, has a master agreement with confidentiality terms, and leadership decides to exclude it before exploring a license of work order history.

Starting state: the eligible scope covers 180,000 work orders from about 400 customers. The university account holds 30% of them. Evidence: count queries from the work order table, grouped by customer, year, job type and closure status. Rule applied: the target use case, a model that triages and schedules preventive and corrective maintenance, needs at least 2,000 closed work orders per job type the buyer names, and no single customer above a quarter of the records.

  • Decision: proceed with a scoped investigation, but drop lab and clean-room equipment from the scope offered, because the remaining 300 orders cannot support the claim.
  • Outcome check: after the leakage search and removal, rerun the counts and confirm zero hits for the university's names, domains and building codes before any sample leaves the company.
  • Counterexample: if the buyer had wanted only HVAC and general building maintenance, the exclusion would have improved fit, because it removes a dominant account without losing any category the buyer needs.
Worked case (illustrative): a facilities maintenance company and its university account
PropertyFull scope (illustrative)Without the university (illustrative)Reading
Closed work orders180,000126,000Volume still well above the use case minimum
Years covered77No year drops out
HVAC preventive maintenance orders41,00033,000Passes
Lab and clean-room equipment orders2,600300Fails: the category almost disappears
Share of records held by the largest remaining customer30%11%Better balance than before
Records still naming the university in free textNot measured1,400Leak: needs removal before any export

What leadership can decide now, and what needs evidence first

Some decisions can be made from the metadata comparison alone; others need evidence that the comparison cannot supply. Keep the two lists separate in the memo so nobody mistakes an assumption for a finding.

  • Decide now: whether the remaining scope clears the volume and history thresholds, which task categories to drop from the offer, and whether to spend time on a scoped investigation at all.
  • Decide now: who owns the exclusion decision (usually the executive sponsor with the account owner) and who signs off on the leakage search.
  • Needs evidence first: whether the excluded customer's contract also restricts records about shared sites, affiliates or subcontracted work, which counsel reads.
  • Needs evidence first: whether other large customers have similar terms, because a second exclusion changes every number in the table.
  • Needs evidence first: the workload and opportunity cost of preparation, which depends on how clean the remaining records are and is estimated from a sample, not from counts.
  • Needs evidence first: whether buyers interested in the use case would accept the narrower scope, which only a scoped conversation answers.

Which record proves the exclusion decision and its outcome?

The proof is an exclusion register entry linked to the queries and counts it was based on, not a status field that says excluded. Each entry names the account, the reason category, the decision owner and date, the exact filter used (customer identifiers, affiliate identifiers, name and domain search terms), the counts before and after, and the result of the leakage search.

Datasheets for Datasets, the documentation framework proposed by Gebru and colleagues, asks whether a dataset is a sample of a larger set, whether it is representative, and, if not, why, including whether instances were withheld. The Hugging Face dataset card template likewise has a place to describe data selection criteria and filtering. Both point to the same practice: the exclusion should appear in the dataset documentation as a selection rule, stated plainly, so a buyer can judge the effect on coverage without learning who the customer is.

The task outcome is verified only when the delivered dataset matches the register. That means rerunning the counts on the final export, confirming the leakage search returns zero hits, and recording the version of the export the check was run on. A ticket closed as done or a message saying the account was removed is not the outcome; the matching counts are.

When excluding the top account makes the dataset easier to license

Exclusion is not always a loss. When one customer dominates the records, a buyer may worry that a model trained on them learns that customer's processes, naming habits and approval rules rather than the industry's. Removing the account can make the remaining data more representative of the market, simplify the rights review, and remove the most identifiable records in one step.

The comparison table tells you which situation you are in. If the categories the buyer needs survive and the concentration measure improves, the exclusion strengthens the offer. If a needed category collapses, the honest options are to narrow the offer, to look for the same category in other customers' records over a longer period, or to ask whether partial use of the excluded account is permitted. Inflating the remaining scope to compensate is not one of them.

How should an exclusion-adjusted scope go into a licensing assessment?

A company that has already run this comparison can describe its offer honestly: what is in, what is out and what the exclusion costs in coverage. SourceX covers qualification, rights review, contracting, preparation and delivery when a business licenses selected records, and an excluded account simply stays outside the scope the company approves and signs before anything is delivered.

Buyers generally expect about two years or more of consistent history and record counts in the tens of thousands within one system, though rarer work can qualify with less, so it is the scope left after the exclusion that gets assessed. Checking whether that scope could qualify begins with a brief assessment of the company, before any records are shared.

Frequently asked questions

Does excluding our largest customer make our data worthless to AI buyers?

Usually not. What matters is whether the remaining records still cover the task types, years and outcomes a buyer needs. If the excluded account held most of one category, that category may have to come out of the offer, but the rest can still qualify.

How do we find every record that belongs to the excluded customer?

Start with the customer field, then search other fields and free text for the customer's names, email domains, site addresses, affiliate names and contract numbers. Records about shared sites or subcontracted work often mention the account without being assigned to it, so the free-text search is not optional.

Can we keep some of an excluded customer's records, such as dates and categories?

Sometimes, if the contract and the reason for the exclusion allow it. Measure that partial scope separately so you know what it adds, and have counsel confirm whether structured fields alone are permitted before relying on it.

Should buyers be told that a customer was excluded?

Buyers should be told that a selection rule removed one account's records and what that did to coverage, without naming the customer. Dataset documentation frameworks such as Datasheets for Datasets expect withheld instances to be described, and stating it up front avoids surprises in the buyer's own review.

What if excluding one customer pushes another above a quarter of the records?

Then the concentration problem has moved rather than gone away. Recompute the table with the new top account in view, and decide whether to cap its share in the offered scope or to accept the concentration and disclose it.

Sources

  • The U.S. Justice Department defines the Herfindahl-Hirschman Index as the sum of the squared market shares of each firm, approaching zero with many similar firms and reaching 10,000 when one firm controls the market. Source
  • SEC Regulation S-K Item 101 asks registrants to describe any dependence on key customers in their business description. Source
  • Datasheets for Datasets asks whether a dataset is a sample of a larger set, whether it is representative, and why not, including whether instances were withheld or unavailable. Source
  • The WILDS benchmark paper reports that standard training yields substantially lower out-of-distribution than in-distribution performance across its real-world datasets. Source
  • The scikit-learn user guide describes GroupKFold as keeping the same group out of both training and test sets, and LeaveOneGroupOut as holding out one group per split. Source
  • The Hugging Face dataset card template includes a data collection and processing section for data selection criteria and filtering methods. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies