Skip to content

AI uses for records

Business size dataset coverage: auditing small and large customers

By SourceX Editorial · Updated

9 sections

Short answer

A business size dataset coverage audit checks whether records from small and large customer accounts support the tasks a model will face. Fix the size definition and its source first, count accounts and complete cases as well as records, then measure the operational traits, such as approval depth, participants and custom terms, that may change difficulty. Stratify by band, keep accounts whole in splits and name the bands the data cannot support.

Checklist

Business size dataset

This page covers business size dataset coverage: auditing small and large customers.Illustrative

Key takeaways

  • Size has no single definition: U.S. SBA standards vary by industry and use employees or receipts, while the EU SME test pairs headcount with turnover or balance sheet, so every size label needs its definition, source and date.
  • The counting basis can flip the picture: when large accounts generate many more records each, record shares and account shares tell opposite coverage stories.
  • Treat size as a pointer and measure the operational traits that plausibly change the task, then test whether difficulty follows the trait or the band.
  • Keep each account inside one split, stratify bands across folds and cluster intervals by account, so a few large customers cannot pass for broad coverage.
  • Name the bands and traits the data cannot support in the coverage statement instead of averaging over them.

What must a business size dataset coverage audit establish?

It must show, for the population the model will serve, how many distinct customer accounts and complete cases each size band contributes, which operational traits differ between bands and whether the task outcome depends on those traits. Record totals alone answer none of these questions.

The audit concerns the customers inside a dataset: the small and large businesses whose orders, tickets, contracts or projects appear in a supplier's records. It is not about the size of the supplier. The difference matters in practice. SourceX's eligibility rule for U.S., Canadian and EU suppliers looks at the supplying company, which needs 50 or more full-time employees at its peak, while the customers recorded in that company's systems can be any size, and their spread is what this audit measures.

The discipline is to avoid two shortcuts: assuming a dataset covers small and large customers because both appear somewhere in it, and assuming large customers mean hard tasks. Both assumptions are testable, and each part of the audit turns one of them into a check.

Small by whose definition? Pin the size label before counting

Write down which definition the size field uses, where each value came from and when it was set, because official definitions disagree. The U.S. Small Business Administration sets size standards industry by industry, generally measured by number of employees or annual receipts (sba.gov, checked 2026-10-27), and its Office of Advocacy uses fewer than 500 employees as its research definition of a small business (2026 FAQ, checked 2026-10-27).

The European Commission's SME definition (checked 2026-10-27) pairs a headcount limit of fewer than 250 staff with either turnover up to 50 million euros or a balance sheet total up to 43 million euros, and a firm inside a larger group may have to count the group's figures. Thresholds also move: in August 2026 the SBA's Office of Advocacy reported proposed rules that would revise size standards for 338 industries and industry groups, which remain proposals until a final rule.

Inside a CRM or billing system, the size field is often a sales segment, a band picked on a sign-up form or a value bought from an enrichment vendor, and it may be years old. Record which of these you have. A segment that sales assigned by deal value is a different variable from headcount, and an audit that mixes the two ends up measuring the sales process.

Then choose band edges for the decision, not for convenience. If the buyer's product is priced by seat, bands drawn on seats beat bands drawn on headcount; if the model will route enterprise escalations, the supplier's own enterprise segment may be the right variable even though it is not a size measure. Whatever you choose, keep the raw value beside the band so a reviewer can re-band later.

Accounts, cases or records: which count decides who looks covered?

Report all three, because the counting basis can reverse the conclusion. National statistics show the effect at scale: the SBA Office of Advocacy's 2026 FAQ counts 36,207,130 small businesses in the United States, which together employ 45.9 percent of private-sector workers. Counted by firm, small businesses dominate; counted by worker, they are under half.

A customer dataset behaves the same way. When a large account files many more tickets or orders than a small one, a record-weighted sample is mostly large-account work and an account-weighted sample is mostly small-account work. Neither is wrong. Each answers a different question, and the intended use decides which share should match the target population.

The restatement uses invented counts for one support archive. Large accounts are 3 percent of accounts but 53 percent of records, so a model evaluated on randomly drawn records is graded mostly on a few dozen customers.

Accounts, cases or records: which count decides who looks covered?
Size band (Illustrative)AccountsComplete casesRecordsWhat a random record draw does
Small, 1-49 employees8206,10031,000Under-weights them: 75 percent of accounts, 22 percent of records
Mid, 50-499 employees2407,90034,000Roughly balanced: 22 percent of accounts, 24 percent of records
Large, 500 or more employees389,80074,000Over-weights them: 3 percent of accounts, 53 percent of records
Total1,09823,800139,000Account and record shares tell opposite stories

Size is a proxy: which operational traits actually change the task?

Measure the traits directly and treat the size band as a pointer to where they might differ. Difficulty in business tasks often comes from structure a record can show: how many approvals a decision needs, how many people touch a case, whether terms are custom, how many systems or entities are involved, how many cases are open at once and how much history the account carries. A 40-person firm with a negotiated contract and three approvers can be harder to serve than a 4,000-person firm on standard terms.

These traits are hypotheses to test, not findings. Read each one from fields the supplier already keeps, such as approval history, participant lists, contract or price-book flags and linked entities, and record its presence per case before any outcome is examined, so the profile cannot be tuned to the answer.

The table lists six traits with an invented share of cases showing each one per band. The pattern to look for is overlap: where many small-account cases carry a trait, size alone cannot stand in for it.

Size is a proxy: which operational traits actually change the task?
Operational traitHow to read it from recordsSmall (Illustrative)Mid (Illustrative)Large (Illustrative)
2+ approversApproval steps logged before the decision closed18%41%63%
3+ participantsDistinct requesters, agents and internal teams on the case24%38%57%
Custom contract termsNegotiated schedule, discount or service level on the account9%27%71%
2+ systems or entitiesIntegrated systems, sites, legal entities or currencies touched12%30%52%
5+ open cases at onceConcurrent open cases for the account when this one opened6%22%68%
Tenure under 1 yearAccount age when the case opened35%21%9%
A matrix of six operational traits against small, mid and large customer bands, with bars showing the illustrative share of cases that carry each trait.
Trait profile by size band. Illustrative: shares are invented to show overlap between bands, not measured data.

The size-band sampling worksheet

Complete the worksheet before drawing a sample or accepting anyone's coverage claim. Each line names what to record, an illustrative entry for the support archive above and the red flag that should pause the audit until it is resolved.

The size-band sampling worksheet
Worksheet lineWhat to recordIllustrative entryRed flag
1. Target populationWhich customers the model will serve, by band and segmentU.S. customers of a mid-market software vendor, all bandsNo target named, so no share can be judged high or low
2. Size definitionDefinition, unit (employees or receipts), source and dateEmployees, from an enrichment vendor, refreshed yearlyA sales segment treated as if it were headcount
3. Size field qualityShare of accounts with a missing, stale or conflicting value6 percent missing; 50 accounts checked against public profilesMissing values dropped or pushed into one band
4. Counting basisAccounts, complete cases and records per band1,098 accounts, 23,800 cases, 139,000 recordsOnly record totals reported
5. Trait profileShare of cases with each operational trait, per bandSix traits from approval, contract and ticket fieldsBand used as the only difficulty variable
6. Minimum per bandFewest distinct accounts and cases a band needs before its results are shownAt least 30 accounts and 300 cases per reported bandA band reported from a handful of accounts
7. Draw and weightsStratified draw by band, and by trait where needed, with weights to the targetEqual cases per band in the test set; results reweighted to the target mixNo weights, so pooled scores mix bands at arbitrary ratios
8. Splits and intervalsGroup key for splits and for clustering intervalsAccounts kept whole, bands stratified across folds, intervals clustered by accountCases from one account in both training and test

Does difficulty follow size, or the traits behind it?

Test it instead of assuming it: compare outcomes across bands, then compare again within each band split by trait. If the gap between small and large accounts disappears once approval depth or custom terms are held fixed, the trait is the variable to stratify on and size was standing in for it.

Three comparisons make the test concrete. First, per band: task success or reviewer-rated difficulty, with intervals. Second, per trait within each band: the same metric for cases with and without the trait. Third, per trait across bands: do small-account and large-account cases with the same trait score alike? Where the third comparison shows no difference, pooling bands for that trait is defensible; where it does, keep the bands separate.

An illustrative reading: suppose large-account cases fail review more often than small-account cases overall, but among cases with two or more approvers the small and large bands fail at about the same rate. Approval depth, not size, is then doing the work, and the test set needs enough multi-approver cases from every band rather than more large accounts.

Keep the statistics honest about the unit. Cases from one account share people, templates and terms, so they are not independent. scikit-learn's cross-validation guide (repository copy, checked 2026-10-27) describes StratifiedGroupKFold as trying to preserve the class distribution in each split while keeping each group within a single split; pass the size band as the stratification label and the account as the group. Anthropic's guidance on evaluation statistics (checked 2026-10-27) recommends clustering standard errors on the sampled unit when items are not independent, which here means the account.

Difficulty itself needs a definition that does not leak the answer. Rate it from intake fields that exist before anyone works the case, as the reward normalization and expert difficulty pages on this site describe, not from how long the case took, which reflects the agent as much as the task.

Could a big record count from large accounts still give enough coverage?

Only if the accounts behind it vary as much as the target population does, and that is the strongest alternative explanation to rule out. In the illustrative archive, 74,000 large-account records come from 38 customers; if a handful of them dominate, the model learns their habits rather than large-account work in general.

Check the spread inside each band: records per account, the top-five account share and results with the largest accounts held out. If removing three accounts changes the large-band score, the band is covered by those accounts and should be described that way.

Selection can bias the small band too. Small customers who left early may have few or no records in a retained archive, and accounts on self-serve plans may raise problems through channels that never reach the ticketing system. Compare the archive's account list with a billing or customer master list for the same period before treating the small band as representative.

In some settings representativeness is also a legal expectation. Article 10(4) of the EU AI Act, which applies to high-risk AI systems, requires data sets to reflect, to the extent the intended purpose requires, the geographical, contextual, behavioural or functional setting in which the system will be used (EUR-Lex, checked 2026-10-27). For other systems it is not a legal duty, but it describes what this audit checks.

Coverage limitations a size audit cannot remove

Some gaps survive the best sampling plan and belong in the coverage statement. Benchmark builders state their strata openly: OpenAI describes the first version of GDPval as spanning 44 occupations selected from the top 9 industries contributing to U.S. GDP, with 1,320 tasks in the full set (checked 2026-10-27). A business dataset should be as explicit about which size bands and traits it covers, and which it does not.

The limitations most likely to remain are listed with the check and the wording each supports.

Coverage limitations a size audit cannot remove
LimitationWhy it happensCheckWording the evidence supports
Too few large accountsLarge customers are few even when their records are manyDistinct accounts per band and the top-five shareLarge-band results come from a stated number of accounts
Proxy or stale size labelsSize comes from a sales segment or an old enrichment valueField provenance and a spot check against an independent sourceBands follow the supplier's segment field, not verified headcount
Departed customers missingRetention rules or plan types keep some accounts out of the archiveArchive accounts compared with the customer master for the periodCovers accounts retained in the system for the window
Traits not recordedApprovals or custom terms live in email or contracts outside the systemShare of cases with each trait field populatedTrait comparisons limited to cases where the field exists
One region or product lineThe supplier sells mostly in one market or productAccounts by region and product within each bandCoverage stated per region and product
Band definitions differThe buyer's size bands are not the supplier'sA crosswalk between both definitionsBands mapped to the buyer's definitions, with losses stated

Asking for size-stratified records in a data request

Name the size bands, the definition and the counting basis in the request itself, along with the traits that matter for your task. Datasets reach buyers through SourceX by request, not from an inventory: the buyer sets out scope, volume, language, format, timing and licensing needs on the request form.

For a size coverage requirement, add the target mix by band, a minimum number of distinct accounts per band, the trait fields to keep, account-level splits and an acceptance test that reports results per band. Any candidate dataset still goes through rights, privacy and quality review, and a request does not imply that matching records exist.

Ask for the size field's raw value, source and refresh date with every account, not just the band, and for the trait fields in the same export as the cases. Without them the buyer cannot rerun this audit, and a coverage claim made by the supplier cannot be checked.

Frequently asked questions

What is business size dataset coverage?

It is how well a dataset represents the customer businesses of different sizes that a model will serve, measured in distinct accounts and complete cases per size band as well as records. A useful coverage report also shows which operational traits differ by band and states the bands the data cannot support.

Which definition of small business should a dataset use?

Use the one that matches the intended use, and record it. U.S. SBA size standards vary by industry and use employees or receipts, the SBA's research definition is fewer than 500 employees, and the EU SME definition pairs fewer than 250 staff with turnover or balance sheet limits. Mixing definitions inside one field makes band comparisons meaningless.

Are large enterprise customers always harder tasks for AI models?

No. Size can go together with approval depth, custom terms or system spread, but those traits also appear at small firms and are absent at some large ones. Measure the traits directly and test whether difficulty follows them within each band before using size as a difficulty label.

Should a dataset match the national mix of small and large businesses?

Only if the model will serve that mix. Match the customers the model will actually face, which for a vendor's support model is its own customer base, and keep weights so pooled results can be restated for a different target.

How do you stop one big customer from dominating a size band?

Count distinct accounts per band, report the top-five account share, cap records per account in the sample and hold out whole accounts in evaluation. If results change when the largest accounts are removed, describe the band as covered by those accounts rather than by the band in general.

Sources

  • The SBA sets small business size standards industry by industry, generally measured by number of employees or annual receipts. Source
  • The SBA Office of Advocacy's 2026 FAQ counts 36,207,130 U.S. small businesses employing 45.9 percent of private-sector workers, using fewer than 500 employees as the definition. Source
  • In August 2026 the SBA proposed revised size standards for 338 industries and industry groups. Source
  • The EU SME definition pairs fewer than 250 staff with a turnover or balance sheet ceiling, and group figures may count. Source
  • Article 10(4) of the EU AI Act requires data sets for high-risk AI systems to take into account the setting in which the system is intended to be used, as far as the intended purpose requires. Source
  • GDPval's first version spans 44 occupations from the top 9 industries contributing to U.S. GDP, with 1,320 tasks in the full set. Source
  • scikit-learn's StratifiedGroupKFold tries to preserve the class distribution in each split while keeping each group within a single split. Source
  • Anthropic recommends clustering standard errors on the sampled unit when evaluation items are not independent. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies