Skip to content

Deal economics

Buyer-ready dataset cost: a five-layer model that charges proof and failures

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A buyer-ready dataset costs far more than the work on its own records. Fully loaded, it carries five layers: records work, the evidence pack buyers and regulators now expect, specialist review, a share of the candidates that failed, and shared overhead. In this article's illustrative model, records work is about a third of 362 loaded hours per accepted dataset. No public benchmark exists, so ask for the model, not a number.

Value drivers

Buyer ready dataset cost

This page covers buyer-ready dataset cost: a five-layer model that charges proof and failures.Illustrative

Key takeaways

  • Price a buyer-ready dataset per accepted dataset in five layers: records work, evidence pack, specialist review, failed-candidate carry and shared overhead.
  • The evidence pack is mostly fixed per dataset, so small datasets carry a larger evidence share than large ones.
  • Failed candidates cost what they consumed before exiting; one late failure at buyer acceptance can outweigh several early rights rejections.
  • Moving provenance checks ahead of records work lowers the loaded cost only when late provenance failures are common.
  • A cost model is credible only if per-dataset hours plus unallocated time reconcile to the hours the team actually worked.

What does a buyer-ready dataset really cost?

More than the work on its own records, usually by a wide margin. A dataset is buyer-ready when it has passed the buyer's acceptance test and ships with the evidence the buyer needs to use it and defend that use: the rights chain, de-identification evidence, documentation and a manifest. Its fully loaded cost adds that evidence work, specialist review, a share of the candidates that never shipped and shared overhead to the hours spent on the records themselves.

There is no public benchmark for this number. Data companies do not publish comparable per-dataset costs, and projects are not measured the same way, so a cost figure quoted without its method tells an investor very little. What can be compared is the structure: which layers a company counts, how it allocates failed work and whether its totals reconcile to the hours its team actually logged.

This article builds that structure as an editable model. Every input is illustrative and expressed in hours, the conversion rates come from public U.S. labor statistics, and nothing here describes SourceX's own costs, prices or margins, which are not public.

What turns a prepared dataset into a buyer-ready one?

Evidence. A prepared dataset is clean and formatted; a buyer-ready dataset also shows where it came from, what rights cover it, how personal details were handled and exactly what it contains. Four public developments make that proof a cost line rather than a courtesy.

The first is documentation norms. The Datasheets for Datasets proposal by Gebru and colleagues, first posted to arXiv in 2018, argues that every dataset should come with a datasheet covering its motivation, composition, collection process and recommended uses, and it organizes its questions into sections that run through preprocessing, distribution and maintenance. Answering them for real business records takes hours from people who know the source systems.

The second is disclosure rules that reach buyers. California's AB 2013, enacted in 2024, requires developers of generative AI systems to post documentation about their training data on or before January 1, 2026, including the sources or owners of the datasets and whether the developer purchased or licensed them. In the EU, the European Commission published a mandatory template on July 24, 2025 for general-purpose AI model providers to summarize their training content, with a data-source section that covers private datasets. A buyer that must disclose needs those facts from its supplier.

The third is privacy evidence. Under HIPAA's expert determination route, 45 CFR 164.514(b)(1) requires the expert who concludes that re-identification risk is very small to document the methods and results that justify the conclusion. Health records are a special case, and outside health care no single rule sets the format, but a written method is the kind of evidence a buyer's privacy reviewer can actually check.

The fourth is the market's track record. In a 2023 arXiv paper covering more than 1,800 text datasets, the Data Provenance Initiative found that widely used dataset hosting sites omitted licenses at rates above 70 percent and recorded them in error at rates above 50 percent. A buyer aware of that gap has reason to check documentation rather than trust a label, and preparing documentation that survives the check is the seller's cost.

The five-layer waterfall: fully loaded hours per accepted dataset

Fully loaded cost per buyer-ready dataset is the sum of five layers, each measured in hours for one accepted dataset. The table gives an illustrative base case with a low-to-high range for every input; replace each figure with logged hours before drawing a conclusion.

In the base case, records work on the shipped dataset, the part most estimates capture, is 120 of 362 loaded hours, about a third. The evidence pack and specialist review add 85 hours that barely move with record count. The failed-candidate carry adds 110 hours that depend almost entirely on how early the pipeline rejects weak candidates. Overhead is applied last at a stated 15 percent, so a reader can see exactly what it contributes.

Two properties matter more than the total. Layers 2 and 3 are largely fixed per dataset, so a dataset half the size does not cost half as much to make buyer-ready. Layer 4 is a property of the pipeline, not of the dataset, so two identical datasets can carry very different loaded costs depending on what failed around them. Setting every input to its low end gives about 156 loaded hours and every input to its high end about 825 (our arithmetic), a fivefold spread.

The five layers, with illustrative hours and each layer's share of the base-case total:

The five-layer waterfall: fully loaded hours per accepted dataset
LayerWhat it coversScales withIllustrative hours: base (range)Share of loaded total
1. Records work on the shipped datasetExtraction, linking, cleaning, de-identification runs and formatting of the accepted recordsRecord count, source systems and linking complexity120 (60-240)33 percent
2. Evidence packDatasheet, data dictionary, provenance and rights chain, de-identification method and results, manifest with counts, acceptance test reportNumber of sources and fields, not rows60 (30-110)17 percent
3. Specialist reviewCounsel review of the rights chain and license scope, privacy review, final quality sign-offNumber of sources and sensitivity of the records25 (12-50)7 percent
4. Failed-candidate carryHours spent on same-cohort candidates that exited before acceptance, divided by accepted datasetsRejection rate and the stage at which candidates exit110 (40-260)30 percent
5. Shared overheadTooling, secure storage, sourcing management and administration, at a stated rate on layers 1-4Rate chosen: 10-25 percent47 at 15 percent (32-79)13 percent
Fully loaded totalHours per buyer-ready datasetAll five layers362 (about 156-825)100 percent
A vertical waterfall of loaded hours per accepted dataset climbing from 120 hours of records work through the evidence pack, specialist review, failed-candidate carry and overhead to 362 hours.
Five layers of fully loaded hours for one buyer-ready dataset. Illustrative: inputs are assumptions, not measured or SourceX data.

How should failed candidates be charged to the datasets that shipped?

Charge each failed candidate the hours it actually consumed before it exited, then divide the cohort's failed hours by the datasets that reached acceptance. Weighting by exit stage matters because a candidate rejected at the rights screen costs a fraction of one that fails the buyer's acceptance test after full preparation.

The illustrative cohort behind layer 4: ten candidates entered detailed work and three reached acceptance. Four exited at the rights screen after about 20 hours each. Two exited during preparation after about 70 hours each, when writing the documentation exposed a provenance gap. One failed buyer acceptance after about 110 hours. That is 330 failed hours, or 110 per accepted dataset, and the single late failure accounts for a third of the carry.

Two allocation choices change the answer and should be stated in any model. Group candidates by the date they entered detailed work, not the period in which they happened to fail, or a quarter with many rejections looks expensive and the next one cheap. And hold candidates still open on the measurement date in their own line rather than counting them as successes or failures, because they resolve later and would distort both.

Five rules keep the carry honest:

  • Count hours, not candidates: an early rejection and a late one are different costs.
  • Group by entry cohort and report the cohort size and the number accepted beside the carry.
  • Hold open candidates in a separate line until they exit or ship.
  • Leave out candidates that never passed a first fit check; their cost belongs to sourcing, a separate measure.
  • Report the carry per accepted dataset and as a share of the loaded total, so a reader can see when it dominates.

Evidence first or evidence last: which order lowers the loaded cost?

Building provenance evidence before records work lowers the loaded cost when late provenance failures are common, and raises it slightly when most candidates fail on rights for other reasons. It is a sequencing decision a data company can test on its own rejection log.

In the evidence-last base case, the two candidates with provenance gaps were found only while the documentation was being written, after 70 hours each. Run a 10-hour provenance check alongside the rights screen instead, before any records work, and those two exit at about 30 hours each. The four rights-screen rejections now absorb the check too, rising to about 30 hours each, while the late acceptance failure stays at 110 because its provenance work was done in either order.

Failed hours fall from 330 to 290, and the carry from 110 to about 97 hours per accepted dataset. Change one input and the case strengthens: if four of the ten candidates had provenance gaps and only two failed the rights screen, evidence-last costs 4 x 70 + 2 x 20 + 110 = 430 failed hours, while evidence-first costs 4 x 30 + 2 x 30 + 110 = 290. The cohort saving grows from 40 hours to 140.

The test for a real company is its own exit data: what share of rejections happen after records work starts, and for what reason. A pipeline that cannot answer has not measured layer 4 at all.

How do you turn these hours into a cost range with public wage data?

Multiply each layer's hours by a loaded hourly rate for the people who do that work, and carry the low and high wage ends through the model rather than a single midpoint. U.S. public statistics give defensible starting ranges; a company's own payroll should replace them.

For records work and evidence preparation, the Bureau of Labor Statistics reports median pay for data scientists of $120,230 a year, or $57.80 an hour, using May 2025 data, with the lowest 10 percent earning less than $67,240 and the highest 10 percent more than $199,130 a year. For rights review, it reports median pay for lawyers of $159,670 a year, or $76.76 an hour, using 2025 data, with the lowest 10 percent under $78,360 and the highest 10 percent over $351,600. Outside counsel bills at its own rates, which a model should enter separately rather than infer from salaries.

Wages are not the full cost of an hour. The bureau's Employer Costs for Employee Compensation release for June 2026 puts wages and salaries at 70.0 percent of private-industry employer compensation costs and benefits at the remaining 30.0 percent. Dividing a wage rate by 0.70 approximates a loaded rate (our arithmetic): the data scientist median becomes about $82.57 an hour and the lawyer median about $109.66.

The wage spread between the 10th and 90th percentiles is roughly threefold for data scientists and more than fourfold for lawyers (our arithmetic), and the hours range is fivefold, so the product spans more than tenfold. That spread is the finding: a single cost-per-dataset figure without its hours, rates and allocation rules cannot be compared with anyone else's.

Does the loaded cost survive a reconciliation test?

Only if the per-dataset costs, plus time no dataset absorbed, add up to the hours the team actually worked in the period. That reconciliation answers the most common objection to any cost model: that a reported deal margin looks healthy until rework, overhead and collection delays are counted.

Run it each quarter. Sum the loaded hours charged to every dataset accepted in the quarter, add the hours held for open candidates and the hours logged to sourcing, sales and administration, and compare the total with timesheet or payroll hours for the same people. A gap of more than a few percent means some work is charged to no dataset, and the per-dataset figure is understated by roughly that share.

Rework hides in the same gap. A dataset prepared twice, because the first delivery failed a check, should carry both passes in layers 1 and 2; a ledger that records only the final pass reconciles only if the first pass turns up somewhere else, usually as unexplained overhead. Collection delays affect cash rather than cost and belong in a payment-timing comparison, such as the delayed-payment guide linked below, not in this waterfall.

What could make this cost model wrong

Four developments could shrink or reshape the loaded cost, and an investor should ask which one a company is most exposed to.

  • Automation could cut the fixed layers. If documentation drafting and provenance checks become largely automated, layers 2 and 3 fall, the small-dataset penalty shrinks and this model overstates the advantage of scale.
  • Buyers could accept less evidence. Licenses were missing at rates above 70 percent on the widely used hosting sites in that 2023 audit, which shows how much AI development has run on poorly documented data. If buyers keep accepting that, the evidence pack is a cost they will not pay for.
  • Disclosure rules could narrow or widen. AB 2013's documentation duty applies to developers of generative AI systems and the EU template to providers of general-purpose AI models; a change in either scope moves layer 2 directly.
  • A low failure rate could mean weak screening. A pipeline that rejects almost nothing has a small layer 4 but may pass problems to the buyer; ask for rejection reasons by stage, not only the rate.

Diligence questions on cost per buyer-ready dataset, SourceX included

Ask for the model and its reconciliation, not a headline number. The same questions apply to any company that sources, prepares or licenses datasets.

Each of the five layers exists in SourceX's process. For businesses licensing operational data to AI developers, it covers qualification and sourcing of the supplier, rights review, contracting, preparation of the data, delivery and payment, and no data is delivered before the supplying company has both approved the scope and terms and signed. It has not published cost-per-dataset results; company and commercial metrics are shared with investors directly through the investor materials request.

Questions to put to any data company:

  • Which of the five layers does your cost per dataset include, and where do the others sit in your accounts?
  • How many candidates entered detailed work in each cohort, how many reached acceptance, and at which stage did the rest exit?
  • What is in your evidence pack, and how many hours does it take for a dataset from one source system versus several?
  • Do you build provenance evidence before or after records work, and what does your rejection log say about that choice?
  • Do loaded hours reconcile to timesheet or payroll hours each quarter, and how large is the unallocated gap?
  • Which loaded hourly rates do you use for preparation and for rights review, and where do they come from?

Methods note: what the five-layer model assumes and leaves out

The five-layer model is an editorial framework, not measured research. All hours, the cohort, the exit stages and the overhead rate are illustrative assumptions chosen to show the mechanics; they are not SourceX figures and do not describe any real company. Public inputs are the Bureau of Labor Statistics pay figures for data scientists (May 2025 data) and lawyers (2025 data), its June 2026 employer compensation release, and the documentation sources named in the text, all checked on October 26, 2026.

Not measured: actual preparation hours for any dataset, rejection rates in any real pipeline, outside counsel rates, tooling and storage prices, rework frequency, and the cost of capital tied up while work waits for payment. The model counts hours from detailed work onward and excludes sourcing outreach before a candidate's first fit check. Ranges are illustrative low and high inputs, not confidence intervals.

Frequently asked questions

Is there a public benchmark for the cost of a buyer-ready dataset?

No. Data companies do not publish comparable per-dataset costs, and projects are measured in different ways, so any single figure needs its method beside it. Compare the structure instead: which cost layers are counted, how failed candidates are allocated and whether totals reconcile to hours worked.

Why charge failed candidates to the datasets that shipped?

Because the accepted datasets are the only ones that can pay for them. Every pipeline rejects candidates, and the hours spent on them are a real cost of producing an accepted dataset. Charging each failure the hours it used before exiting, divided across accepted datasets from the same cohort, keeps that cost visible.

What belongs in the evidence pack for a licensed dataset?

Typically a datasheet or dataset card, a data dictionary, the provenance and rights chain, the de-identification method and results, a manifest with counts, and the acceptance test report. Buyers that must describe training sources under rules such as California's AB 2013 or the EU training-content template need facts from this pack.

Do smaller datasets cost less to make buyer-ready?

Less in total, but more per record. The evidence pack and specialist review scale with the number of sources and fields rather than rows, so a dataset half the size keeps most of those hours. That fixed share is why small datasets carry a larger evidence burden.

Does SourceX publish its cost per dataset?

No. SourceX has not published cost, price or margin figures, and the model in this article uses illustrative inputs only. Company and commercial metrics are shared with investors directly through the investor materials request at sourcex.si/investors.

Sources

  • The Data Provenance Initiative's 2023 audit of more than 1,800 text datasets reported license omission above 70 percent and error rates above 50 percent on widely used dataset hosting sites. Source
  • Datasheets for Datasets proposes that every dataset come with a datasheet covering motivation, composition, collection process and recommended uses, with question sections through distribution and maintenance. Source
  • California AB 2013 (2024) requires generative AI developers to post training data documentation on or before January 1, 2026, including dataset sources or owners and whether datasets were purchased or licensed. Source
  • On July 24, 2025 the European Commission presented a mandatory template for general-purpose AI model providers to summarize training content, with a data-source section covering private datasets. Source
  • Under 45 CFR 164.514(b)(1), the expert determination method requires the expert to document the methods and results of the analysis that justify the finding of very small re-identification risk. Source
  • BLS: data scientists' median pay was $120,230 a year ($57.80 an hour) in May 2025; lowest 10 percent under $67,240, highest 10 percent over $199,130. Source
  • BLS: lawyers' median pay was $159,670 a year ($76.76 an hour) in 2025; lowest 10 percent under $78,360, highest 10 percent over $351,600. Source
  • BLS Employer Costs for Employee Compensation, June 2026: wages and salaries were 70.0 percent and benefits 30.0 percent of private-industry employer compensation costs. 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