Skip to content

AI data market

Serviceable enterprise data supply: five filters on official counts

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Serviceable enterprise data supply is the number of companies in a defined population whose records pass system, rights, language, quality and buyer-fit filters. Official tables count 306,000 EU enterprises and 48,113 Canadian businesses with 50 or more people; confirmed U.S. counts cover only the 21,041 firms with 500 or more. Illustrative filter ranges then leave about 1,300 to 49,000 companies, so measure rights and buyer fit first.

Value drivers

Serviceable enterprise data

This page covers serviceable enterprise data supply: five filters on official counts.Illustrative

Key takeaways

  • Count companies before records: a licensable dataset needs a company that can sign, and official business statistics count companies by size.
  • Eurostat's preliminary 2024 figures give 251,000 medium and 55,000 large EU enterprises; ISED's 2025 edition gives 48,113 Canadian employer businesses with 50 or more employees.
  • The confirmed U.S. counts split only at 500 employees, 21,041 firms above and 6,374,594 small employers below, so U.S. firms with 50 to 499 stay out of the frame until a finer table is read.
  • With labeled illustrative filters, the serviceable count spans 1,337 to 48,952 companies, about 37 times; rights clearance (3.00x) and buyer fit (2.67x) move it most.
  • An end-to-end screen of 400 randomly drawn companies would narrow the joint pass rate to an interval a few points wide, which no amount of desk research can do.

What does serviceable enterprise data supply measure?

It measures how many companies in a defined population could deliver at least one record set that a buyer can license, prepare and use in a stated period. It is the supply-side companion to a demand model built from buyer budgets, and it answers a separate diligence question: is the supply a company says it can reach real, countable and large enough once the filters that actually remove companies have run?

The unit is the company, not the byte or the record, because a company is what signs. Rights review, contracting, preparation and payment happen once per supplier, and official statistics count businesses by size class, so a company frame can be tied to published tables in a way that a count of records or terabytes cannot.

A useful answer needs three things: a population defined by a public table with its year and size definition, a short list of filters each tied to a kind of evidence, and the arithmetic shown so a reader can swap any input. Everything below is built that way. SourceX publishes no supplier, pipeline or market-size figures, and nothing on this page is one.

Where do official business counts stop? The 50-person cut in three regions

They stop at different places: official tables count companies at a 50-person cut in the EU and Canada, but the confirmed U.S. counts split only at 500. The cut mirrors the size focus SourceX publishes, companies whose full-time staff reached 50 or more at their peak; any other cut works the same way if a table splits there.

Eurostat's preliminary 2024 structural business statistics, published on 9 December 2025, count 251,000 medium-sized enterprises (50 to 249 persons employed) and 55,000 large ones (more than 249) in the EU business economy, 306,000 together. Innovation, Science and Economic Development Canada's Key Small Business Statistics 2025 counts 27,780 employer businesses with 50 to 99 employees, 16,953 with 100 to 499 and 3,380 with 500 or more, 48,113 together, with the size totals dated December 2024.

The U.S. is the gap. The SBA Office of Advocacy, which defines a small business as one with fewer than 500 employees, counts 21,041 large businesses among 6,395,635 employer firms in its February 2026 FAQ, built on the Census Bureau's 2022 Statistics of U.S. Businesses, which leaves 6,374,594 small employer firms. The Census Bureau's standard size bands for those statistics break at 20, 100 and 500 employees, not at 50, and no table read here splits the small employers at 50, so the frame counts only the 21,041 U.S. firms with 500 or more employees and leaves out those with 50 to 499.

Each count comes with a definition, a date and a blind spot:

Where do official business counts stop? The 50-person cut in three regions
Region and sourceSize definition and dateCount at 50 or moreWhat it does not measure
EU: Eurostat structural business statisticsPersons employed; preliminary 2024 data, published 9 December 2025306,000 (251,000 medium plus 55,000 large)Peak headcount, full-time status, or enterprises outside the business economy
Canada: ISED Key Small Business Statistics 2025Employees, counted at enterprise level; size totals dated December 202448,113 (27,780 plus 16,953 plus 3,380)Peak headcount; comparison with earlier editions, which counted differently
U.S.: SBA Office of Advocacy, from 2022 Census business statisticsEmployees; 2022 data in the February 2026 FAQ21,041 (500 or more only)The 50 to 499 band, which no table read here isolates
Frame: the three combined, illustrativeMixed definitions and years375,154U.S. firms with 50 to 499 employees, companies that shrank below 50 after their peak, and companies in other regions

Five removal steps, each resting on a different kind of evidence

Five filters remove most of the frame, and each rests on a different kind of evidence. Their order matters for measurement, not for arithmetic: in a multiplicative model the product is the same whichever filter runs first.

They run here from the one with a public anchor to the ones that need fieldwork:

Checklist · 5 items

  1. F1 System of record with multi-year history. Eurostat's 2023 enterprise survey found ERP software in 66% of medium and 86.3% of large EU enterprises, so the low and high values are anchored for the EU and assumed elsewhere. ERP use is a proxy: a qualifying record family can also sit in a help desk or CRM.
  2. F2 Rights clear for at least one record family. Customer contracts, privacy notices, platform terms and earlier AI training licenses can each remove a record set. No public table measures it; SourceX's legal framework page describes checking terms source by source and excluding data or narrowing scope when a use is restricted.
  3. F3 Language buyers can use. SourceX's FAQ accepts non-English data from U.S., Canadian and EU companies and asks companies in several other countries for at least 90% English; for a buyer the test is whether the records' languages fit its task. No public rate exists.
  4. F4 Quality: history, volume and outcomes. As a general guide, buyers look for years of consistent history and tens of thousands of records in one system, with rare work the exception, and records need the outcome of the work, not only the request. No public rate exists.
  5. F5 Buyer fit. The record family has to match documented demand. SourceX's investor page says its focus areas, software, BPO and contact centers, professional services, finance, healthcare administration and logistics, are matched against documented demand from AI labs and model developers. No public rate exists for any population.

0 of 5 ticked. Tick each item as you go.

Frame times five shares: the arithmetic, line by line

Multiply the frame by the five filter shares at their low values and again at their high values; the two products bound an illustrative serviceable count. Every input is labeled anchored or illustrative, and none is a SourceX figure.

Low: 375,154 x 0.66 x 0.20 x 0.60 x 0.30 x 0.15 = 1,337 companies. High: 375,154 x 0.863 x 0.60 x 0.90 x 0.70 x 0.40 = 48,952 companies. The high end is about 37 times the low end, which is the honest width of a supply estimate built from one anchored filter and four assumed ones:

Frame times five shares: the arithmetic, line by line
InputLowHighLabel and source
Frame: companies with 50 or more people375,154375,154Anchored: Eurostat, ISED and SBA Office of Advocacy counts; mixed definitions; U.S. at 500 or more only
F1 System of record with history66%86.3%Anchored: Eurostat 2023 ERP use in medium and large EU enterprises
F2 Rights clear for one record family20%60%Illustrative assumption
F3 Language buyers can use60%90%Illustrative assumption
F4 Quality: history, volume, outcomes30%70%Illustrative assumption
F5 Buyer fit with documented demand15%40%Illustrative assumption
Serviceable companies1,33748,952Illustrative result, not an estimate of SourceX supply

Which uncertainty is a reading problem, and which needs fieldwork?

The U.S. population gap is a reading problem; rights and buyer fit need fieldwork, and between them they move the result most. Rights clearance has the widest swing, closely followed by buyer fit. In a multiplicative model each input's influence is its high value divided by its low value, whatever the other inputs are: rights 3.00 times, buyer fit 2.67, quality 2.33, language 1.50 and system of record 1.31. The frame has no range here, because U.S. firms with 50 to 499 employees are left out rather than bracketed; counting them could only raise both ends.

The two kinds of uncertainty call for different work. The population gap closes at a desk: any table that splits U.S. firms at 50 employees, whether an official tabulation or a business register, settles it. Rights and buyer fit are measurement problems that desk research cannot close, because no public source records how often companies can license a record family or whether it matches what buyers ask for.

That is where diligence time pays. A supply claim that cites a large company count but no measured rights or fit rate has done only the cheap part of the work.

Ranked horizontal bars of each input's high value divided by its low value in an illustrative supply model, rights 3.00, buyer fit 2.67, quality 2.33, language 1.50 and system of record 1.31, beside a panel giving the illustrative result of 1,337 to 48,952 companies.
Illustrative: company counts from Eurostat, ISED and the SBA Office of Advocacy and ERP shares from Eurostat; the other filter ranges are assumptions.

An end-to-end screen of 400 companies: the sampling plan

Draw companies at random from the frame, screen each one through all five steps, and estimate the joint pass rate directly. Measuring each filter on a separate sample and multiplying assumes the filters are independent, and they probably are not: a company with clean contracts may also keep better records.

Size the sample for the interval you need. For a proportion near 50%, 385 companies give a 95% margin of about 5 points; for the small joint rates this model implies, a Wilson interval behaves sensibly near zero where the textbook formula does not. The table shows what 400 screened companies would yield at four hypothetical outcomes, none of them a result, and the companies each interval implies on the 375,154 the tables count for certain. Stratify by region and size band, and report refusals, because companies that decline a screen are not a random subset:

An end-to-end screen of 400 companies: the sampling plan
Passed all five filtersObserved rate95% Wilson intervalCompanies on the 375,154 frame
4 of 4001.0%0.4% to 2.5%1,501 to 9,379
8 of 4002.0%1.0% to 3.9%3,752 to 14,631
20 of 4005.0%3.3% to 7.6%12,380 to 28,512
40 of 40010.0%7.4% to 13.3%27,761 to 49,895

Do AI budgets, data stocks or funding rounds measure supply?

No. Each measures something upstream or downstream of a company that can sign. AI spending measures buyer capacity, much of it compute; global data volume measures stored bytes, most of which nobody may license; venture funding measures investor appetite for the category. None counts companies whose records clear rights, language, quality and fit, so none can stand in for the frame or the filters.

Company counts overstate too, in a quieter way. The frame is a ceiling on companies, not on willing suppliers: a company that passes all five filters may still decline to license, and that conversion step is something a company in this market should evidence with its own funnel, not with national tables.

Supply also means little without demand. A large serviceable pool in record families no buyer asks for is inventory, not a market, so read a supply model beside a demand model held to the same unit discipline, such as the published methodology for data licensing market reports linked below.

What would push the true count outside the range?

Seven things could, and each one pushes the count in a stated direction:

At a glance

What would push the true count outside the range

  • The uncounted U.S. band: firms with 50 to 499 employees are left out of the frame, so the true count is higher by whatever share of the 6,374,594 U.S. small employer firms reaches 50 employees and passes the filters.
  • Definitions that do not match: Eurostat counts persons employed, ISED and the U.S. statistics count employees, and none records peak headcount, so firms that once passed 50 and shrank are missing while firms padded by part-time staff are included.
  • Editions that do not compare: ISED notes that its 2025 counts are calculated at the enterprise level and are not comparable with earlier editions, so a trend built across editions would mix methods.
  • Correlated filters: if rights, quality and fit move together, the true joint rate falls outside the product of independent ranges, in either direction.
  • Supply already spoken for: SourceX asks suppliers to confirm their data has not been licensed for AI training before, so every earlier deal elsewhere removes records from the pool faster than a company count shows.
  • Shifting buyer fit: synthetic data or new model capabilities could narrow the record families buyers want, cutting F5 without changing any count.
  • Regions outside the frame: companies in several countries beyond the U.S., Canada and the EU, from the UK and India to Latin America, can qualify when at least 90% of their data is in English, so the frame understates reachable supply there.

Diligence questions for any claimed supplier pool

The useful questions ask for the frame, the filters and the denominators, not the headline. A supply claim that cannot name its population table, its filter evidence and the share of screened companies that passed is a slogan, however large the number.

SourceX can be read through the same lens. Its investor page describes licensing business data directly from companies and licensing it onward to AI developers in its own name, managing sourcing, rights review, contracting, preparation, delivery and payment, with focus areas matched against documented demand from AI labs and model developers. It publishes no supplier, pipeline or conversion figures, and the page says it shares company and commercial metrics with investors directly. A reader weighing SourceX itself as an investment should know that its offering runs under Rule 506(c), which limits purchasers to verified accredited investors, and that the investor page takes requests for materials.

Six questions test any supply claim, including one from SourceX:

  • Which population table, year and size definition define your frame?
  • How many companies did you screen, how were they drawn, and how many passed each filter?
  • Is your rights pass rate measured per company or per record family, and who decided each case?
  • Which record families match documented buyer demand today, and how is that demand documented?
  • How many qualified companies declined to license, and why?
  • How much of your pool has already licensed the same records to someone else?

Methods note: tables read, rounding rules and gaps

Built on 2026-10-28. Counts come from Eurostat's news release of 9 December 2025 (preliminary 2024 structural business statistics), ISED's Key Small Business Statistics 2025, the SBA Office of Advocacy's FAQ of February 2026 (from the 2022 Statistics of U.S. Businesses), the Census Bureau's description of its size bands, and Eurostat's E-business integration statistics for 2023 ERP use. Each was read as a search excerpt on the publisher's own domain because direct fetches are blocked from the build environment, so a reviewer should open every page before release.

Arithmetic: EU 251,000 + 55,000 = 306,000; Canada 27,780 + 16,953 + 3,380 = 48,113; U.S. 21,041 (500 or more), with 6,395,635 - 21,041 = 6,374,594 small employer firms left out because no table read here splits them at 50; frame 306,000 + 48,113 + 21,041 = 375,154 at both ends. The serviceable range multiplies the frame by all five filter shares at their low and at their high values and rounds once, half up, to whole companies. Swing ratios are high divided by low for each input, rounded half up to two decimals. Wilson intervals use z = 1.96 and are rounded half up to one decimal, and company counts apply the rounded percentages shown and round half up to whole companies. The script is supply_model_v2.py.

Not measured: U.S. firms with 50 to 499 employees; any filter rate except the ERP anchor; correlation between filters; companies' willingness to license; demand or prices; regions outside the EU, Canada and the U.S.; and anything about SourceX's own suppliers, applicants or pipeline, none of which is public.

Frequently asked questions

How is serviceable supply different from a data licensing TAM?

A TAM sizes what buyers spend; serviceable supply counts the companies that could deliver a licensable record set. They rest on different evidence, buyer budgets and seller disclosures on one side and business counts and screened pass rates on the other, and neither can be derived from the other.

Why use 50 people as the size cut?

It mirrors the size focus SourceX publishes and lines up with the medium and large classes Eurostat and ISED report. Any cut works if the table behind it splits there; the U.S. bands read here break at 20, 100 and 500, not at 50, which is why the frame counts U.S. firms only at 500 or more.

Can the ERP adoption rate stand in for having good data?

Only partly. ERP use shows that a company runs a system of record, which is why it anchors the first filter, but rights, language, quality and buyer fit are separate tests that ERP adoption says nothing about.

How many companies would a screening study need?

About 385 give a 5-point margin on a rate near 50%, and around 400 is a practical size for the small joint rates this model implies, read with a Wilson interval. Stratify by region and size band and report refusals, since companies that decline bias the rate.

Does SourceX publish its own supply or pipeline numbers?

No. SourceX publishes no supplier, pipeline, buyer or market-size figures and shares company and commercial metrics with investors directly through its investor page. The ranges here are illustrative and are not SourceX estimates.

Sources

  • Eurostat's preliminary 2024 structural business statistics (news of 9 December 2025) count 251,000 medium-sized enterprises (50 to 249 persons employed) and 55,000 large enterprises in the EU business economy. Source
  • ISED's Key Small Business Statistics 2025 counts 27,780 employer businesses with 50 to 99 employees, 16,953 medium (100 to 499) and 3,380 large (500 or more) as of December 2024, calculated at enterprise level and not comparable with earlier editions. Source
  • The SBA Office of Advocacy's February 2026 FAQ, based on the 2022 Statistics of U.S. Businesses, counts 21,041 large businesses and 6,395,635 employer firms and defines a small business as having fewer than 500 employees. Source
  • The Census Bureau says tabulations of its business statistics by employment size use less than 20, 20 to 99, 100 to 499, less than 500 and 500 or more. Source
  • Eurostat's E-business integration statistics report ERP software use in 2023 by 66% of medium and 86.3% of large EU enterprises. Source
  • SourceX's how-it-works page says it focuses on companies with 50 or more full-time employees at peak and that suppliers confirm the data has not been licensed for AI training before. Source
  • SourceX's FAQ accepts non-English data from U.S., Canadian and EU companies and asks companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia for at least 90% English. Source
  • SourceX's investor page describes direct licensing from companies and onward licensing to AI developers, focus areas matched against documented demand, and sharing company and commercial metrics with investors directly. Source
  • SourceX's legal framework page says each source tool's terms are checked during review and that data is excluded or the deal scoped to fit when terms restrict a use. 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