Skip to content

Deal economics

Data or labor? How to spot a data startup services business model

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A data startup has a services business model when most of its fulfillment hours start and stop with each buyer, whatever its deck says. Code every hour by dataset version and by buyer: hours spent once per version behave like an asset, hours repeated per buyer behave like staffing. The second-buyer hours ratio, plus expense disclosures arriving with ASU 2024-03 and IFRS 18, shows which business you are buying.

Value drivers

Data startup services

This page covers data or labor? How to spot a data startup services business model.Illustrative

Key takeaways

  • Labor is not the tell; the driver of labor is. Hours that scale with datasets fit licensing, hours that scale with buyers or delivered units are services.
  • Filings show three cost patterns: research labor that maintains one shared asset, royalties that rise with every sale, and crowd annotation that follows each client's projects.
  • The second-buyer hours ratio, hours the next buyer of the same version adds divided by the first buyer's hours, separates the models better than revenue labels or blended margins.
  • In an illustrative model, three businesses spend 420 hours on their first buyer; by the fifth buyer one needs 100 hours each and another still needs 388.
  • ASU 2024-03 and IFRS 18 will expose employee compensation inside expense lines from 2027, giving investors a public yardstick for private companies.

What makes a data startup a services business in disguise?

A data startup is a services business when most of its fulfillment hours begin and end with one buyer. The label on the pitch deck, whether data platform, marketplace or licensing company, does not change that. What changes the economics is whether the next buyer of the same dataset needs a fresh round of work or receives something that already exists.

That gives investors a narrower, testable question than the usual debate about labels: for each hour of fulfillment labor, what made it necessary? Some hours are spent once per dataset version, on rights review, extraction, de-identification and base quality checks, and every later buyer shares them. Other hours repeat for every buyer, on labels written to that buyer's taxonomy, collection to its protocol, format conversion and acceptance support. The first kind behaves like an asset. The second kind behaves like a staffing contract.

Labor itself is not the tell. Almost every data company employs people who touch the data, and some of the most durable data products depend on research staff. The tell is the driver: hours that scale with the number of datasets are compatible with licensing economics, while hours that scale with the number of buyers or delivered units are services economics, and should be valued on labor margin, utilization and buyer retention.

Three cost patterns that public filings already show

Public filings show three distinct patterns, and only one of them is services. Reading what a company says sits inside its cost line, and what makes that cost grow, separates them more reliably than revenue labels do.

ZoomInfo, a subscription data company, says in its 10-K for 2024 that cost of service includes direct expenses related to the support and operations of its services and research teams, including salaries and benefits, alongside costs such as third-party hosting, third-party data and amortized internally developed software. That is labor, but in a subscription model research work maintains the shared database all subscribers draw on, so the hours attach to the asset rather than to any one customer.

Getty Images says in its 10-K for 2025 that its cost of revenue consists primarily of royalties owed to content contributors, and its quarterly report for the period ended September 30, 2024 says it expects cost of revenue to trend higher in absolute dollars as revenue grows. Cost that rises with every sale is not automatically services: here it is a revenue share paid to the people who made the content, with little labor repeated per license.

TELUS International, in a 2021 prospectus filed with the SEC, described data annotation delivered through crowdsourced contractors and named, as a risk, that its business would suffer if those individuals were classified as employees rather than independent contractors. Annotation volume follows each client's projects, so the labor is per buyer and per delivered unit. That is the services pattern, and the classification risk is part of its cost structure. The table summarizes the three patterns and one label that spans several:

Three cost patterns that public filings already show
Company and documentWhat the disclosure showsWhat drives the costReads as
ZoomInfo, 10-K for 2024Cost of service includes services and research teams' salaries and benefits, hosting, third-party data and amortized softwareKeeping one database current for all subscribersLabor per asset: compatible with licensing economics
Getty Images, 10-K for 2025Cost of revenue is primarily royalties owed to content contributorsEach license soldA revenue share that scales with sales, not repeated labor
TELUS International, 2021 prospectusData annotation through crowdsourced contractors; reclassification as employees named as a riskEach client project's volumeLabor per buyer and per unit: services economics
Labelbox, November 2024 announcementA labeling platform, managed labeling services and an expert hiring marketplace under one brandWhich of the three a customer buysOne label, three economics: split before valuing

The hours ledger: code every fulfillment hour twice

The hours ledger tags every fulfillment hour with two codes, the dataset version it served and the buyer it served, with all buyers as a permitted value. Once hours carry both codes, the services share stops being a matter of opinion and becomes a sum anyone can recompute from time entries and contractor invoices.

At a glance

The hours ledger: code every fulfillment hour twice

Ask for the ledger at the level of steps, because each step has a default driver that the evidence either confirms or overturns. Rights review can repeat when a new buyer's use needs fresh approval, and a refresh is per asset only if every buyer receives the same refresh. Score the step by what the records show, not by the default. The steps, their default drivers and the record that settles each one:

  • Rights review and supplier approval of a dataset version: per asset by default; per buyer when each new use needs new approval. Settled by approval records dated against each delivery.
  • Extraction from the source systems: per asset. Settled by extraction job logs keyed to the dataset version.
  • De-identification and privacy checks: per asset, unless a buyer asks for extra fields that need fresh review. Settled by review tickets naming the fields.
  • Labeling to the company's own taxonomy: per asset. Settled by labeling tasks that carry no buyer code.
  • Labeling, collection or annotation to a buyer's taxonomy or protocol: per buyer. Settled by the buyer's guideline document and the tasks that cite it.
  • Format conversion, schema mapping and sample cuts for one buyer: per buyer. Settled by delivery tickets.
  • Delivery, acceptance support and buyer questions: per buyer. Settled by support logs and acceptance correspondence.
  • Scheduled refreshes: per asset when one refresh serves every licensee; per buyer when each refresh is cut to one buyer's specification. Settled by refresh manifests listing recipients.

Business-model comparison with scope-normalized metrics

The comparison below normalizes each model to the scope a buyer actually receives, one dataset version or one delivered unit, so a services company cannot look like a licensing company by quoting revenue per employee or a blended gross margin. The decisive column is the second-buyer hours ratio: the hours the second buyer of the same version adds, divided by the hours the first buyer took including all the shared work.

The ratio bands are editorial rules of thumb for this article, not industry standards. Below about 0.2, an additional buyer is close to pure licensing. Between 0.2 and 0.5, there is a licensing core with a services tail that should be priced and margined separately. Above 0.5, each sale is mostly a new job, and the business should be valued like a services firm whatever it calls itself. The models, with their scope-normalized metrics:

Business-model comparison with scope-normalized metrics
ModelHours that repeat for each buyerScope-normalized metric to requestSecond-buyer hours ratio (illustrative band)Value it as
Catalog licenseDelivery, format conversion, light supportPer-buyer hours per dataset versionUnder 0.2Licensing: price per buyer, renewals, buyer count per version
Configured licenseBuyer filters, schema mapping, custom refresh schedulesPer-buyer hours reported apart from base-version hours0.2 to 0.5Licensing core plus a services tail priced on its own
Managed data serviceCollection, labeling and review to the buyer's specificationHours per delivered unit, by buyerAbove 0.5Services: labor margin, utilization, retention of the largest buyer
Expert staffing marketplaceMatching, vetting and supervising people for each engagementSupervision hours per hour of expert work soldAbove 0.5Staffing: durability of the spread between billed and paid labor
Data software platformOnboarding and support; the customer's own staff do the labelingSupport hours per customer account after onboardingUnder 0.2 once onboardedSoftware: retention and expansion per account
Royalty-bearing content licenseLittle labor; a royalty owed on each saleRoyalty and delivery cost per licenseLow in hours, but cost per sale persistsLicensing with a revenue share that caps gross margin

One dataset version, five buyers: an illustrative hours curve

In the illustrative model, three businesses each spend 420 fulfillment hours to serve their first buyer, yet their average hours per buyer diverge sharply by the fifth. The numbers are invented to show the mechanism and describe no real company; the script that produces them is kept for reviewers.

The catalog licensor spends 400 hours once per dataset version and 20 hours per buyer. With five buyers on the same version, average fulfillment falls from 420 to 100 hours per buyer, a 76 percent drop, and its second-buyer hours ratio is about 0.05. The configured licensor spends 300 hours once and 120 per buyer: it falls from 420 to 180 hours, a 57 percent drop, with a ratio of about 0.29. The managed data service spends 40 hours once and 380 per buyer: it falls only from 420 to 388 hours, an 8 percent drop, with a ratio of about 0.9.

Two readings follow. First, a revenue chart cannot tell these businesses apart in their first year, because one buyer per dataset looks identical in all three; the difference appears only when the same version sells again. Second, customer-level concentration means different things in each. In the managed service, 98 percent of the hours at five buyers are buyer-specific, so losing the largest buyer removes revenue and frees staff at once. In the catalog licensor, losing a buyer removes revenue but leaves the 400 shared hours sunk in an asset that can still sell.

Line chart of average fulfillment hours per buyer as one dataset version sells to one through five buyers: a catalog license falls from 420 to 100 hours, a configured license from 420 to 180, and a managed data service only from 420 to 388.
Illustrative: invented inputs from the article's hours model, not company data.

Does a platform label hide services economics?

Sometimes, and the way to find out is to split revenue and hours by offering before comparing anything. The strongest alternative explanation for a services-heavy profile is that the company really is a software or licensing business that happens to sell some services, and the label is fair. The opposite case is also common: a platform name sits on top of revenue that is mostly managed labor.

Labelbox shows why the split matters. In a November 13, 2024 announcement, it presented three offerings: its platform for teams that label in house, managed labeling services that combine its own project managers with expert labelers, and Alignerr Connect, a marketplace for hiring expert AI trainers into a customer's own processes. Under one brand sit software economics, services economics and staffing economics, each deserving its own multiple.

So the test for a label is mechanical. Ask what share of revenue came from each offering, then run the hours ledger on each share. If the platform line carries small per-account support hours while the managed line carries most of the growth, the company is growing as a services firm, and the platform is a delivery tool for that labor rather than the product investors are paying for.

New expense disclosures will put labor on the face of the filings

Comparable public evidence on labor intensity is about to improve, which gives investors a reference point for private companies. The FASB's ASU 2024-03 requires public business entities to disaggregate relevant expense captions in the notes into categories that include employee compensation, purchases of inventory and depreciation, depletion and amortization. FASB's summary defines employee compensation broadly, including cash consideration, share-based payment and postemployment benefits.

Timing is set by ASU 2025-01, which clarified that public business entities first apply the disclosures in the first annual reporting period beginning after December 15, 2026, with interim periods following within annual periods beginning after December 15, 2027. For a calendar-year company, the first full disclosure arrives with its 2027 annual report.

IFRS reporters are on a similar track. IFRS 18, which replaces IAS 1 and applies from 1 January 2027, requires a company that presents operating expenses by function to disclose the amounts of employee benefits, depreciation, amortisation and two other specified expenses included in each line item. The IASB's effects analysis says preparers that now disclose only totals will need to gather more detailed information.

For an investor, the practical step is to ask private data companies for the same split now: employee compensation and contractor labor inside cost of revenue, by dataset version and by buyer. A company that cannot produce it does not yet know which business it is running.

What could make the hours test give the wrong answer

The hours test fails in predictable ways, and each failure has a check that a diligence team can run before relying on the result:

  • Shared-hour allocation. Hours spent on several buyers at once have to be split, and the rule can move the ratio. Ask for the rule, then recompute two dataset versions by hand.
  • Capitalized labor. Labor booked into internally developed software or a database asset leaves cost of revenue today and returns later as amortization, which ZoomInfo's cost of service definition includes. Add capitalized hours back before computing ratios.
  • Contractor classification. Crowd or freelance labor can sit outside headcount while driving most fulfillment cost. TELUS International named reclassification of annotators as employees as a risk to its business, which shows that the labor line can change for legal rather than operational reasons.
  • Missing time records. Many young companies log hours by person, not by dataset version and buyer. Treat estimates reconstructed after the fact as weaker evidence and say so in the memo.
  • Automation trends. Tooling can turn per-buyer hours into per-asset hours over time, so a high ratio today is not a verdict for life. Measure the ratio by cohort of dataset versions and look for a falling line.
  • Royalties and data costs. A business can have few repeated hours yet a large cost per sale, as Getty Images' royalty-driven cost of revenue shows. A low hours ratio does not by itself prove high margins.

Questions for any data company about its labor, SourceX included

SourceX helps companies turn selected business records into licensed AI datasets and handles sourcing and qualification, rights review, contracting, data preparation, delivery and payment along the way. Preparation is fulfillment labor, so the hours ledger applies to SourceX as it does to any company in this market.

Two published facts shape where its hours should fall. A supplying company approves the scope and terms and signs before any data reaches SourceX, after which SourceX may use, sublicense and distribute the de-identified dataset as that agreement allows, which is what would allow shared preparation to serve more than one buyer where the agreement permits. Datasets for AI buyers are sourced on request rather than held in stock, so an investor should ask how often a request reuses an existing dataset version and how often it starts new work. SourceX does not publish hours, ratios or margins; it shares company and commercial metrics with investors directly, and materials for its raise from accredited investors under Rule 506(c) are available on request through its investor page. Questions to put to SourceX or any data company:

  • What share of last year's fulfillment hours carried a buyer code, and what share served all buyers of a dataset version?
  • What is the second-buyer hours ratio, by cohort of dataset versions, and is it falling?
  • How many dataset versions have been delivered to two or more unrelated buyers?
  • Inside cost of revenue, how much is employee compensation, how much is contractor labor and how much is royalties or data costs?
  • How much labor was capitalized, and what would the ratio be with it added back?
  • What share of buyer-specific hours went to the largest buyer?

Methods note: sources, model inputs and what was not measured

The hours ledger, the step defaults, the second-buyer hours ratio, its bands and the comparison table are editorial frameworks built for this article on 27 October 2026. Public evidence comes from SEC filings by ZoomInfo, Getty Images and TELUS International, a Labelbox announcement, FASB's summary of ASU 2024-03 and ASU 2025-01, and the IASB's effects analysis for IFRS 18, each checked on its publisher's domain that day through search excerpts, which reviewers should confirm against the full documents.

The illustrative curve comes from a short deterministic script: average hours per buyer equals shared hours plus buyers times per-buyer hours, divided by buyers, and the second-buyer hours ratio equals per-buyer hours divided by shared plus per-buyer hours. Inputs are 400 and 20 hours, 300 and 120 hours, and 40 and 380 hours. No randomness is involved.

Not measured: no company's time records were coded, no filing's labor share was computed, and the three public cases each illustrate one cost pattern rather than a complete ledger. No SourceX data is used, and no SourceX results are reported.

Frequently asked questions

What is a services business model in a data startup?

It is a data company whose fulfillment work mostly repeats for each buyer: collection, labeling or preparation done to that buyer's specification and paid for as it is delivered. Revenue grows with staff and contractors rather than with the number of datasets owned. It can be a sound business, but it should be valued on labor margin and buyer retention rather than on licensing multiples.

Is a large labor share in cost of revenue always a bad sign?

No. A data subscription company can carry research staff in cost of service because those people maintain one database that every subscriber uses. What matters is whether the labor repeats for each new buyer. Labor attached to a shared asset can support licensing economics, while labor attached to each buyer cannot.

How do you calculate a second-buyer hours ratio?

Take the fulfillment hours logged to the second buyer of a dataset version and divide them by all hours spent on that version up to and including the first buyer's delivery. Compute it for each version and report the median by cohort. It needs time entries coded by dataset version and by buyer, so ask whether those codes exist before asking for the number.

When will public companies show employee compensation inside cost of revenue?

Under ASU 2024-03, as clarified by ASU 2025-01, public business entities apply the expense disaggregation disclosures from the first annual period beginning after December 15, 2026. IFRS 18, effective 1 January 2027, requires companies presenting expenses by function to disclose employee benefits in each line item. Calendar-year companies will show the split in their 2027 annual reports.

Can a services-heavy data company become a licensing business?

It can if it keeps the rights to what it builds and sells the same dataset versions again, which turns per-buyer hours into shared hours. The evidence is a falling second-buyer hours ratio across cohorts and a growing count of versions delivered to unrelated buyers. Promises of a future catalog are not evidence until those second sales happen.

Sources

  • ZoomInfo's 10-K for 2024 says cost of service includes direct expenses of its services and research teams, including salaries and benefits, with costs such as hosting, third-party data and amortized internally developed software. Source
  • Getty Images' 10-K for 2025 says its cost of revenue consists primarily of royalties owed to content contributors. Source
  • Getty Images' quarterly report for the period ended September 30, 2024 says it expects cost of revenue to trend higher in absolute dollars as revenue grows. Source
  • A 2021 TELUS International prospectus says its business would be harmed if individuals providing data annotation through its crowdsourcing solutions were classified as employees rather than independent contractors. Source
  • Labelbox announced on November 13, 2024 a platform for in-house teams, managed labeling services combining its project managers with expert labelers, and Alignerr Connect, a marketplace to hire expert AI trainers. Source
  • FASB's summary of ASU 2024-03 lists employee compensation, purchases of inventory and depreciation, depletion and amortization among required expense categories, with employee compensation covering cash, share-based and postemployment benefits. Source
  • ASU 2025-01 states that public business entities first adopt the ASU 2024-03 disclosures in the first annual reporting period beginning after December 15, 2026, with interim periods within annual periods beginning after December 15, 2027. Source
  • The IASB's effects analysis says IFRS 18 replaces IAS 1 from 1 January 2027 and that companies presenting expenses by function will need more detailed information to disclose five specified expenses, including employee benefits, for each line item. 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