Skip to content

Leadership and readiness

Company size and dataset value diligence: test records, not headcount

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Headcount is a weak proxy for dataset value: it counts labor, while buyers pay for recorded decisions with outcomes, history and licensable rights. Use company size only as an eligibility floor, then test six multipliers that turn staff into usable records: recording roles, records per person-year, retained history, outcome links, licensable share and buyer fit. In an illustrative case, a 70-person field-service firm out-yields a 700-person manufacturer about thirty to one.

Value drivers

Company size dataset value

This page covers company size and dataset value diligence: test records, not headcount.Illustrative

Key takeaways

  • Headcount measures labor input; dataset value depends on what work was recorded, kept, linked to outcomes and cleared for licensing.
  • Buyers that screen by size pair the floor with other tests: micro1's program page asks for 30 or more full-time employees plus mature operations and substantial historical data.
  • The headcount-to-records bridge multiplies six factors, and one factor near zero erases any advantage of size.
  • Replace the org chart with records: a role-to-system map, counts by year, retention and migration history, outcome completeness and a seeded sample.
  • A signed license or a polished sample proves neither coverage nor quality; measure both on records you draw yourself.

Does a bigger company hold a more valuable dataset?

Not reliably. Headcount tells you how many people a company paid; dataset value depends on how much of their work was recorded in a system, kept long enough to show patterns, linked to an outcome and cleared for outside use. Two companies with the same headcount can sit far apart on every one of those factors.

That is why company size dataset value diligence should treat headcount as a screen, not a measure. Buyers of business data already use it that way. micro1's Company Data Partnerships page asks for established companies with 30 or more full-time employees, mature operations and substantial historical operational data, and it ties what it pays to factors such as dataset size, workflow complexity, data quality, the domain expertise represented and how unusual the operating knowledge is. Headcount opens the door, but none of those value factors is headcount.

The investor version of the question is sharper. When a data company describes its supply in staff numbers, for example suppliers with 500 or more employees or a pipeline of companies employing tens of thousands of people, it is describing the population it screened, not the records it can deliver. This article converts headcount into usable records with a six-factor bridge, works one illustration, lists the evidence to request and names the ways the test can mislead.

Why headcount became the default screen, and what it actually counts

Headcount is used because it is easy to find, not because it predicts records. Since the SEC's 2020 amendments to Regulation S-K, Item 101(c) has required public companies to describe their human capital resources, including the number of persons they employ. No comparable rule asks any company to disclose how many decision records it holds, how far back they reach or who may license them.

The number itself is softer than it looks. In 2023 the SEC's Investor Advisory Committee recommended that companies break their headcount into full-time, part-time and contingent workers, which says something about how much a single total hides. For a private company it is worse: estimates built from public profiles mix contractors, part-time staff and former employees, and they show this year rather than the peak.

Headcount still carries some signal. Peak headcount bounds how much work a company could ever have recorded, and a floor screens out firms too small to hold years of consistent records in one system. SourceX applies it in exactly that limited role. Its floor is a peak of 50 or more full-time employees, contractors excluded, for companies in the U.S., Canada and the EU; past that screen, its general guide is that buyers want a consistent record that typically spans two years or more, holding tens of thousands of records or more within a single system, with each deal assessed on its own. The floor decides who gets assessed; the records decide what the data is worth.

The headcount-to-records bridge: six multipliers between staff and value

Usable records equal recording staff, times records per person-year, times years retained, times the share linked to an outcome, times the licensable share, times the share that fits a written buyer task. Headcount enters only through the first factor, and each later factor can fall close to zero.

The bridge multiplies on purpose. A company with a large staff and a short retention window, or with long history and no outcome fields, earns no partial credit for size, because a near-zero factor anywhere erases it. Ask for the evidence behind each factor separately: a company that is strong on one tends to assume it is strong on all.

Each row names the factor, what it measures, the record that shows it and the way it usually fails.

The headcount-to-records bridge: six multipliers between staff and value
MultiplierWhat it measuresRecord that shows itHow it fails
Recording rolesShare of staff whose work produces a record in a systemRole-to-system map: each job family and the systems it writes toLine, field or floor work done on paper, radio or memory
Records per person-yearHow many distinct decisions each recording role logsRecord counts by system, year and creating roleTemplate entries and auto-generated rows counted as decisions
Years retainedHistory still retrievable with its fields intactRetention settings and the migration history of each systemA system replacement that kept balances but dropped case history
Outcome linkageShare of records tied to what happened nextCompleteness of status, resolution, payment or rework fields by yearCases closed without a result, or outcomes kept elsewhere with no join key
Licensable shareShare the company may license after rights exclusionsExclusion ledger for customer contracts, third-party content and opt-outsRecords held for customers, or built around material the company does not own
Buyer fitShare matching a named task and acceptance testWritten buyer requests mapped to record typesA large stock that no stated task needs

Worked example: 700 employees against 70

In this illustration the smaller company ends up with about thirty times as many buyer-usable records. Both companies are invented, and every multiplier is an assumption chosen to show the mechanism, not a measurement of any real firm or industry.

Company A is a 700-person manufacturer. Most of its people work on the line, where decisions never reach a system, so about 90 staff write records: quality, planning and service teams logging nonconformance reports and work orders. An ERP replacement in 2023 kept three years of history. Company B is a 70-person home-services firm whose dispatchers, customer service staff and technicians log every job in a field-service system, with nine years of history and an invoice or callback attached to most jobs.

The headcount ranking reverses: about 5,700 usable records for A against about 175,000 for B. The ratio matters less than where A lost ground. Recording roles and retention, not size, decided the result, which is why the evidence list in the next section asks for those records first.

Worked example: 700 employees against 70
Step (Illustrative)Company A: 700 staffCompany B: 70 staffWhat drove the gap
Recording staff9058Line work at A leaves no record
Records per person-year3001,000Every job at B is a ticket
Years retained39The ERP replacement at A dropped history
Linked to an outcome35%80%B records invoices and callbacks
Licensable share40%70%Customer drawings and NDAs at A
Fits a written buyer task50%60%Similar demand assumptions
Buyer-usable recordsAbout 5,700About 175,000Headcount ranking reversed
A slope chart linking headcount to buyer-usable records for two illustrative companies: the 700-person manufacturer falls to about 5,700 usable records while the 70-person field-service firm rises to about 175,000, so the lines cross.
Illustrative: headcount rank against usable-record rank for the worked example. Values are assumptions, not measured data.

Which records prove workflow coverage and record quality?

Eight records rebuild the bridge from evidence, and an investor should ask for them before discussing headcount at all. A company that knows its data can produce most of them from system administration screens, policy files and its own exclusion log without sending a single customer record.

Retention deserves its own question because legal minimums are short. The IRS tells businesses to keep most tax records for three years, employment tax records for at least four, and records supporting a bad debt or worthless securities claim for seven. A company that kept only what tax rules require may hold far less history than its age suggests, whatever its size.

The table below is the dataset diligence evidence request list for this question. Each request names the headcount assumption it replaces, the test you run and the red flag.

Which records prove workflow coverage and record quality?
RequestReplaces this assumptionTest you runRed flag
Role-to-system mapMore staff means more recorded workCount roles that write to a system of record and compare with total staffMost staff write nothing a system keeps
Record counts by system, year and creating roleVolume scales with sizePlot counts per year and look for gaps, spikes and automated rowsCounts include system-generated or duplicated rows
Retention settings and migration historyAn older, larger company has more historyMatch each system's start date and purge rules to the years offeredA migration that kept balances but not case history
Outcome field completeness by yearRecords carry their own resultsMeasure the share of closed records with a resolution, payment or rework valueOutcomes held in email or a finance system with no join key
Duplicate and template shareEach record is a separate decisionHash text fields and group near-identical entriesA few templates account for most rows
Exclusion ledger with countsThe company may license what it holdsReconcile exclusions for customer contracts, third-party content and opt-outsExclusions described as policy, with no counts
Seeded stratified sample protocolA sample shows the archiveDraw your own sample from a manifest fixed in advanceOnly a curated sample is offered
Written buyer requests mapped to record typesAny large archive has buyersMatch each request's task and acceptance test to the record types offeredNo request names a task the records serve

How to sample records instead of trusting the org chart

Fix a manifest first, then draw a seeded, stratified sample that qualified domain reviewers score without seeing the company's own quality labels. The manifest lists every system, table, date range and record count in scope; once it is fixed, nobody can swap better records in after the draw.

Stratify by system, year, creating role and outcome status, and oversample the oldest years and anything that crossed a migration, since history and fields break there. Datasheets for Datasets, the documentation framework from Gebru and colleagues, asks the same of any dataset: whether it holds every instance or a sample, what the larger set is, whether the sample represents it and how that was checked.

Match reviewers to the record type. A claims examiner reads adjuster notes, a maintenance planner reads work orders and counsel reads the exclusion ledger; a generalist scoring everything will mistake fluent records for useful ones. Record the seed, the strata, the sample size in each and every exception with its denominator, so a second reviewer can reproduce the result.

Score each sampled record on four questions: does it record a decision rather than a status change, does it carry the context a model would need, is an outcome attached, and could the company license it. The pass rates become measured multipliers that replace the assumptions in the bridge.

Does a signed agreement or a polished sample settle the question?

No. A signed license shows what the company agreed to let someone use; it says nothing about whether the records cover the workflow a buyer needs or are good enough to train or test on. A polished sample shows only what the company chose to show.

Buyers measure value against tasks, not against the size of the firm behind the records. OpenAI built GDPval from 1,320 tasks across 44 occupations in nine sectors, each based on actual work product from professionals with an average of 14 years of experience, and says it recruited experts from different practice areas and firms of different sizes to make the benchmark representative. To that kind of buyer, a dataset matters when it covers a task family from start to outcome.

Documentation is no shortcut either. The Data Provenance Initiative, auditing 1,858 public finetuning datasets, reported license omission of 70 percent or more and miscategorization error rates of 50 percent or more on widely used hosting sites. Labels on curated public collections drift that far, so a data room summary of a private archive needs testing against records, however large the company that wrote it.

What could make the bridge mislead an investor

The bridge can mislead in five ways, and each calls for a further check rather than a return to headcount.

  • Size and records can move together inside one narrow segment. Among firms running the same software for the same workflow, headcount may track volume closely and the bridge adds cost without changing the ranking; test that on a few companies before dropping it.
  • The multipliers are not independent. Companies with long retention often keep better outcome fields too, so multiplying separate estimates can exaggerate gaps; where possible, measure the combined pass rate on the same sampled records.
  • Demand can move. micro1 said in October 2026 that, over the next 12 months, it intends to put $1 billion into acquiring and licensing enterprises' operational data for reinforcement learning environments. That plan is stated, not spent, and buyer priorities could shift toward tasks where history depth or volume matters less.
  • Small suppliers can be narrow. A 70-person firm's records may come from a few large customers or one region, so numerous records can still cover a thin slice of the work; check customer and site concentration before treating volume as coverage.
  • Rights can change after the review. Notices, retention rules and customer contracts evolve, and a licensable share measured today can shrink when a key contract renews with new restrictions.

Questions to ask any company that screens suppliers by size, SourceX included

Ask how a company gets from its size screen to usable records, using definitions you set. The same questions apply to SourceX.

SourceX carries selected business records from a company through qualification, rights review, contracting, preparation and delivery as a licensed AI dataset, and the company's approval of the scope and terms, plus its signature, comes before any delivery of data. Its published headcount floor, described above, is an eligibility screen; whether a company's records have value is decided on the records.

SourceX has not published supplier, record or commercial metrics, including any conversion from headcount to usable records; it hands company and commercial metrics to investors directly instead. Its raise is limited to accredited investors under Rule 506(c), and anyone evaluating the company can ask for investor materials through the investors page.

Checklist · 6 items

  1. Of the companies that passed your size floor, what share reached a signed license, and what stopped the rest?
  2. For your last ten datasets, what were recording staff, years retained and outcome completeness, measured on records rather than estimated?
  3. Which multiplier most often caused a rejection: retention, outcomes, rights or buyer fit?
  4. Can we draw our own seeded sample from a fixed manifest for a dataset you consider typical?
  5. How do you treat contractors and outsourced teams whose work appears in a supplier's systems?
  6. Have smaller suppliers ever met a buyer's acceptance test better than larger ones, and what explained it?

0 of 6 ticked. Tick each item as you go.

Methods note: how the bridge and request list were built

The bridge, the worked example and the eight-request list are editorial frameworks built for this article on 2026-10-27. Evidence comes from micro1's program page and its October 2026 announcement, the SEC's 2020 Regulation S-K release and its Investor Advisory Committee's 2023 recommendation, the IRS's record retention guidance, OpenAI's GDPval, Datasheets for Datasets and the Data Provenance Initiative audit, each checked that day. Companies A and B are invented, and every multiplier in the example is an assumption.

Not measured: the actual relationship between headcount and usable records in any industry, typical values for any multiplier, and SourceX's own supply. A study that measured them would fix a sampling frame of assessed companies in advance, record each multiplier on sampled records with reviewer agreement, report denominators and intervals, and name the companies that most contradict the average. That study has not been run, and no results are published.

Frequently asked questions

Is employee count a good proxy for how much usable data a company has?

Only as a floor. Headcount counts people, while usable data depends on which roles record their work, how long records are kept, whether outcomes are attached and what the company may license. An illustrative 70-person field-service firm can out-yield a 700-person manufacturer, so ask for record counts by system and year instead.

Why do AI data buyers set minimum employee counts?

A floor is cheap to check and screens out firms unlikely to hold years of consistent records in one system. micro1's program asks for 30 or more full-time employees alongside mature operations and substantial historical data, and SourceX's floor is 50 or more full-time employees at peak, contractors excluded. Both then judge the records themselves.

What evidence shows workflow coverage in a company's records?

A role-to-system map, record counts by system and year, retention and migration history, and outcome field completeness show coverage from the company's own files. A seeded sample drawn from a fixed manifest and scored by domain reviewers then confirms it. Together they show which steps of a workflow were recorded and kept.

How do contractors and outsourced teams affect dataset value?

Their work may be recorded in the company's systems, in a vendor's systems or nowhere. Headcount floors usually leave contractors out, yet records they create inside the company's systems can still be valuable if contracts allow licensing. Check where each team's records sit and who controls them before counting or discounting them.

Does SourceX publish how many usable records its suppliers hold?

No. SourceX has not published supplier, record or commercial metrics; an investor gets company and commercial metrics directly from SourceX, can request materials through sourcex.si/investors and can ask how any metric is defined, with any offering limited to verified accredited investors.

Sources

  • micro1's Company Data Partnerships page asks for established companies with 30 or more full-time employees, mature business operations and substantial historical operational data, and ties compensation to factors including dataset size, workflow complexity, data quality, domain expertise and uniqueness of operational knowledge. Source
  • micro1 announced in October 2026 a stated plan to spend $1 billion over the next 12 months acquiring and licensing enterprise operational data, used to build reinforcement learning environments. Source
  • The SEC's 2020 amendments to Regulation S-K (Release 33-10825) require a description of a registrant's human capital resources under Item 101(c), including the number of persons employed. Source
  • The SEC Investor Advisory Committee's September 2023 human capital recommendation proposed breaking headcount disclosure into full-time, part-time and contingent workers. Source
  • The IRS's general rule is to keep records three years, employment tax records at least four years, and records for bad debt or worthless securities claims seven years. Source
  • GDPval comprises 1,320 tasks across 44 occupations in nine sectors, built from actual work product of professionals averaging 14 years of experience, with experts recruited across practice areas and firm sizes for representativeness. Source
  • Datasheets for Datasets asks whether a dataset contains all possible instances or a sample, what the larger set is, whether the sample is representative and how representativeness was validated. Source
  • The Data Provenance Initiative audit of 1,858 finetuning datasets in 44 collections found license omission of 70 percent or more and error rates of 50 percent or more on popular dataset hosting sites. 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