Skip to content

Deal economics

Data preparation economies of scale: a four-layer cost curve to test

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Data preparation economies of scale are real but bounded. Setup and oversight costs spread thinly as one source grows, so cost per record falls fast, then flattens at a floor set by exception handling. In this article's illustrative model, cost per 1,000 accepted records falls 52 percent from 25,000 to 1,000,000 records, yet 1,000,000 records from 40 sources cost 1.9 times as much.

Value drivers

Data preparation economies

This page covers data preparation economies of scale: a four-layer cost curve to test.Illustrative

Key takeaways

  • Preparation cost has four layers that scale differently: automated passes, exception handling, oversight and per-source setup.
  • Only oversight and setup spread with volume; exception handling sets a floor that more records alone do not lower.
  • In the illustrative model, the full cost is almost nine times an automation-only estimate, and exceptions are 70 percent of cost at 1,000,000 records.
  • Scale counted in records misleads when the records come from many small sources, because each source restarts setup and quality sampling.
  • Ask for cost per accepted record by layer, source size and quarter, with rework and dropped records included, before crediting a scale story.

Does data preparation get cheaper at scale?

Yes, until it reaches a floor. The cost of preparing one more record falls quickly while fixed work (connecting a source, mapping its schema, sampling its quality, documenting the dataset) is spread across more records, then flattens at the cost of handling the records automation cannot finish. Whether a data business enjoys data preparation economies of scale depends less on how many records it processes than on how large its sources are and how many exceptions its records throw.

That distinction decides an investment case. A pitch that extrapolates the steep early part of the curve will overstate future margins, while a diligence team that looks only at the floor will miss real operating leverage. The evidence that settles it is cost per accepted record split into layers, by source size and over time, with dropped records and rework included.

This article builds that split as a transparent model. Every input is labeled illustrative and expressed in reviewer-minutes, never money; public research supplies the mechanisms, not the numbers. SourceX's own costs and volumes are not public, and nothing here describes them.

Four cost layers that respond to growth differently

Preparation cost splits into four layers, and each answers to a different driver: automated passes grow with records, exceptions grow with records times difficulty, oversight grows with deliveries and datasets, and setup grows with sources. Only the last two shrink per record as volume rises, and setup shrinks only when the volume comes from the same source.

Read the third column first: a layer that is fixed per source but large per record falls only when sources get bigger, not when the business gets busier. The public evidence for each layer's behavior:

Four cost layers that respond to growth differently
LayerWhat drives itScales withPublic evidence
Automated passExtraction, de-identification and model labeling run on every recordRecords; cost per record keeps falling as tools improveWang and colleagues, 2021: GPT-3 labels matched human-label results at 50 to 96 percent lower labeling cost
Exception handlingRecords the automated pass flags or cannot finish, resolved by a personRecords times exception rate times minutes per exceptionSegment Anything, 2023: time per mask fell from 34 to 14 seconds, then rose back to 34 on harder objects
OversightQuality sampling, review of machine flags and dataset documentationDeliveries and datasets, not recordsNorthcutt and colleagues, 2021: reviewers confirmed 51 percent of algorithmically flagged label errors
Source setupConnecting a system, mapping its schema, scoping rights and configuring redactionSources; it repeats for every new systemGoogle Research, 2021: data cascades often begin early, at data definition and collection

What automation removes, and what it leaves for people

Automation removes most of the per-record work on easy records and leaves a residue that is harder per item than the average record was. That residue, plus the people needed to check the machine, is why the per-record curve stops falling.

The gains are large and documented. Wang and colleagues reported in 2021 that models trained on GPT-3 labels matched the performance of models trained on human labels at 50 to 96 percent lower labeling cost, and that a mixed strategy, in which people relabel the GPT-3 labels with the lowest confidence scores, beat either labeler alone. The exception queue is built into the best result.

The residue is harder. In the Segment Anything data engine, model assistance cut the time annotators needed per mask from 34 to 14 seconds during the assisted-manual stage. In the semi-automatic stage that followed, with confident masks filled in by the model, average time per remaining mask went back up to 34 seconds because those objects were harder to label (Kirillov and colleagues, 2023). The tool kept improving while the work left for people got harder.

Checking the machine is a cost of its own. When Northcutt and colleagues used an algorithm to flag likely label errors in ten widely used test sets, human reviewers confirmed 51 percent of the flags on average, so the other half were false alarms someone still had to read. Presidio's documentation says that because it relies on automated detection, there is no guarantee it will find all sensitive information, and that additional systems and protections should be used. Oversight shrinks per record as volume grows, but it does not reach zero.

The illustrative cost curve from 25,000 to 1,000,000 records

In the model, cost per 1,000 accepted records falls from 290.1 reviewer-minutes at 25,000 records to 139.7 at 1,000,000 when one source grows, a 52 percent drop, and most of it comes early: 40 percent by 100,000 records, then another 20 percent from there to 1,000,000. The floor is 135.9 minutes, the per-record work that no amount of volume removes; at 1,000,000 records, exception handling, including exceptions on records later dropped, makes up 70 percent of the single-source cost.

The second scenario reaches the same volumes by adding sources of 25,000 records each and combining them into one dataset. Setup, abandoned setups and quality sampling repeat for every source, so cost per 1,000 accepted records only moves from 290.1 to 266.7. At 1,000,000 records the many-source route costs 1.9 times the single-source route, with identical tools and staff.

That gap is the investor's point. Record counts are the easiest scale metric to report and the least informative; the median source size, and the share of accepted records that came from large sources, say far more about where a company sits on this curve. The values below are illustrative model outputs, not measurements:

The illustrative cost curve from 25,000 to 1,000,000 records
Accepted recordsOne growing source, minutes per 1,00025,000-record sources, minutes per 1,000Many-source premium
25,000290.1290.11.0x
100,000174.4272.11.6x
250,000151.3268.51.8x
1,000,000139.7266.71.9x
A line chart of reviewer-minutes per 1,000 accepted records against volume: one growing source falls from 290.1 to 139.7 near a 135.9 floor, while 25,000-record sources stay near 266.7.
Cost per 1,000 accepted records, one growing source versus many 25,000-record sources. Illustrative model values, not measured data.

From the automation-only estimate to the full cost: an editable waterfall

At 100,000 accepted records from one source, the full illustrative cost is 174.4 reviewer-minutes per 1,000 accepted records, almost nine times the 20 minutes an automation-only estimate would show. Each line adds a cost that a tool demo or a single-deal figure tends to leave out.

The inputs, all illustrative: an automated pass and its monitoring at 0.02 reviewer-minutes per record; exceptions on 6 percent of records at 1.5 minutes each; rework on 1 percent of records at 1.5 minutes; 8 percent of prepared records dropped at acceptance; a quality sample of 457 records per source reviewed at one minute each; 600 minutes of governance and documentation per dataset; 2,400 minutes of setup per source; and one in four source setups abandoned halfway, which adds 400 minutes for every source that delivers. Replace each with a company's logged values.

The quality sample is the one line with a formula behind it. Estimating an error rate expected near 5 percent to within 2 points at 95 percent confidence takes about 457 records whether the source holds 25,000 records or 1,000,000, so the share of records reviewed falls from 1.8 percent to 0.05 percent across that range. The waterfall, in reviewer-minutes per 1,000 accepted records:

From the automation-only estimate to the full cost: an editable waterfall
LineMinutesRunning totalWhat moves it
1. Automated pass and monitoring20.020.0Tool time per record; the figure an automation pitch quotes
2. Exception handling90.0110.0Exception rate times minutes per exception
3. Rework after quality review15.0125.0Share of records sent back and the time to redo them
4. Work on records dropped at acceptance10.9135.9Share of prepared records that never count
5. Quality sampling4.6140.4Sample size per source divided by records per source
6. Governance and documentation6.0146.4Fixed work per dataset divided by its size
7. Source setup24.0170.4Setup per source divided by records per source
8. Setup on sources that never delivered4.0174.4Abandonment rate and how far setup got

Halving the exception rate does not halve the cost

Not when the exceptions that remain are harder. If better automation halves the exception rate from 6 to 3 percent but each remaining exception takes 2.5 minutes instead of 1.5, exception minutes fall from 90 to 75 per 1,000 records, a sixth rather than a half, and the cost floor moves from 135.9 to 119.6 minutes instead of to 87.0.

That is the Segment Anything pattern expressed as arithmetic. Automation takes the confident cases first, and what reaches a person is the ambiguous record, the odd layout, the free-text note with a name buried in it. A company forecasting margin from a falling exception rate should show minutes per exception beside it; the rate falling while minutes per exception rise is normal, and the rate falling with minutes flat is the claim that needs evidence.

The break-even condition is short: when the exception rate halves, exception cost falls only if minutes per exception less than double. Ask for both series by quarter and by record type, because a drift toward easier record types can flatter both at once.

Scale, source mix or bookkeeping: which one explains the margin?

A falling cost per record can come from genuine scale, from a shift toward larger and cleaner sources, or from costs that moved out of the reported line, and only the first one lasts. Test the other two before crediting scale.

Source mix is the commonest impostor. A company whose recent quarters were dominated by one large, well-structured source will show the single-source curve, and its next quarters, built from many small systems, can show the other one. Cost per accepted record by source size band exposes the difference that an aggregate hides.

Bookkeeping is the second. Rework booked as a separate project, abandoned setups written off as business development, sampling done by a partner and documentation deferred until a buyer asks all lower a reported preparation cost without changing the work. Timing matters too: a cost curve in minutes says nothing about when a buyer pays, and a delivery that is accepted but unpaid still consumed every minute above.

These questions separate the three explanations:

  • What is cost per accepted record for sources under 25,000 records, from 25,000 to 250,000, and above 250,000, by quarter?
  • What share of preparation minutes went to exceptions, and how did minutes per exception move as the exception rate fell?
  • Where are rework, abandoned setups and dataset documentation recorded, and are they inside the cost per record you report?
  • How many records were prepared and then dropped at acceptance, and is their cost charged to the datasets that shipped?
  • Which preparation work does a processing partner carry, and what happens to the curve if that work comes in house?

What could bend the curve the wrong way

Five developments could flatten or raise a preparation cost curve, or shrink the value of being low on it. Each has a public marker an investor can track:

  • Fixed documentation could grow per dataset. For high-risk AI systems, Article 10 of the EU AI Act requires training, validation and testing data to follow data governance and management practices that cover data-preparation operations such as annotation, labelling, cleaning, updating, enrichment and aggregation. Work that is fixed per dataset does not spread across a portfolio of small datasets.
  • Automation could commoditize the falling part of the curve. If model labeling keeps cutting cost on easy records, as it did in Wang and colleagues' 2021 results, rivals reach the same floor and buyers can capture the saving in price.
  • Data problems could surface late. In the Data Cascades study, built on interviews with 53 practitioners, data cascades were pervasive, with 92 percent prevalence, and often invisible and delayed (Sambasivan and colleagues, 2021). Late discovery turns into rework after delivery, the most expensive place to find a defect.
  • Residual privacy risk could demand more human review, not less. Automated detectors make no completeness promise, as Presidio's documentation states, so buyers can ask for larger samples or full review of sensitive record types, pushing those records back up the curve.
  • Demand could stay concentrated. Reddit's Form 10-K for fiscal 2025 states that, to date, "substantially all of the contract value associated with our licensing revenue is derived from two of our partners." A low cost per record helps little when a few buyers set the price.

Diligence questions for a data preparation business, SourceX included

Ask for the layers, not the average. A company that can rebuild its cost per accepted record from logged minutes by layer, source size and quarter can defend a scale claim; one that quotes a single blended figure cannot yet. Put the same questions to SourceX.

SourceX is one application of the question. Data preparation is one of six jobs it takes on for businesses licensing operational data to AI developers, alongside sourcing and qualification, rights review, contracting, delivery and payment, and de-identification is done either by SourceX itself or by its processing partner. Rights scoping comes ahead of any record-level preparation, because a supplying company delivers nothing to SourceX until it has signed the agreement on a scope and terms it has approved.

SourceX has not published preparation costs, volumes or margins, and this article presents none; investors get its company and commercial metrics directly. Its raise is limited to accredited investors under Rule 506(c), and materials can be requested on its investor page, the place to bring the questions below:

Checklist · 5 items

  1. What is your cost per accepted record by layer: automated pass, exceptions, oversight and setup?
  2. How large is your median source, and what share of accepted records came from sources above 100,000 records?
  3. How have the exception rate and minutes per exception moved over the last four quarters, by record type?
  4. Which preparation steps does a partner perform, and how is that cost counted in your figures?
  5. Which documentation duties apply per dataset under the rules your buyers face, and who carries them?

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

Methods note: the cost model's inputs and blind spots

The four-layer model, the two scale curves, the waterfall and the exception sensitivity were built for this article on 2026-10-26. Every input is illustrative and chosen to show the arithmetic, not drawn from SourceX or any company: 0.02 reviewer-minutes per record for the automated pass, a 6 percent exception rate at 1.5 minutes, 1 percent rework at 1.5 minutes, 8 percent of prepared records dropped, a 457-record quality sample per source, 600 minutes of documentation per dataset, 2,400 minutes of setup per source and one in four setups abandoned halfway. The sample size uses the normal approximation for a proportion: 1.96 squared, times 0.05, times 0.95, divided by 0.02 squared, rounded up.

Public sources supply mechanisms only: Wang and colleagues (2021) on model labeling cost, the Segment Anything paper (2023) on per-mask annotation time, Northcutt and colleagues (2021) on flag precision, Presidio's documentation on the limits of automated detection, the Data Cascades paper and Google Research's 2021 blog post, Article 10 of the EU AI Act as consolidated on EUR-Lex, and Reddit's Form 10-K for fiscal 2025.

Not measured: any company's actual minutes per layer, how setup varies by source system, whether setup gets cheaper as similar systems repeat, compute cost, wage differences across locations and when buyers pay. The model holds every input constant across scale, which real operations will not, so treat it as a template for a company's own logs rather than an estimate of any business.

Frequently asked questions

What are economies of scale in data preparation?

They are the fall in cost per prepared record as volume grows, mostly because fixed work such as connecting a source, mapping its schema, sampling quality and documenting a dataset is spread over more records. The fall slows once per-record exception handling dominates, so the gains are largest early and within a single large source.

Does automation push data labeling cost toward zero?

No. Automation cuts cost on easy records, sometimes sharply, but leaves exceptions that are harder per item and creates flags that people must check. In the Segment Anything data engine, time per remaining mask rose back to 34 seconds once the model handled the easier objects.

Why does the number of sources matter more than the number of records?

Because setup and quality sampling repeat for every source system. In this article's illustrative model, 1,000,000 records from 40 sources of 25,000 records each cost 1.9 times as much per record as 1,000,000 records from one source.

How should an investor read a data company's gross margin?

Ask what cost of revenue includes: exceptions, rework, dropped records, abandoned setups and documentation, or only tooling and direct labeling. Then ask for it by source size band and by quarter, because a margin earned on one large, clean source does not carry over to a portfolio of small ones.

Has SourceX published its preparation costs or volumes?

No. SourceX has not published preparation costs, volumes or margins, and this article uses illustrative inputs only. It shares company and commercial metrics with investors directly, on request through its investor page.

Sources

  • Wang and colleagues (2021) found that models trained on GPT-3 labels matched those trained on human labels at 50 to 96 percent lower labeling cost, and that having people relabel the lowest-confidence GPT-3 labels beat either labeler alone. Source
  • The Segment Anything paper (2023) reports per-mask annotation time falling from 34 to 14 seconds in the assisted-manual stage and rising back to 34 seconds in the semi-automatic stage because the remaining objects were harder to label. Source
  • Northcutt and colleagues found that human reviewers confirmed 51 percent of algorithmically flagged label-error candidates on average across ten widely used test sets. Source
  • Presidio's documentation states that because it uses automated detection there is no guarantee it will find all sensitive information, and that additional systems and protections should be employed. Source
  • Sambasivan and colleagues (CHI 2021) interviewed 53 AI practitioners and found data cascades pervasive (92 percent prevalence), invisible and delayed, but often avoidable. Source
  • Google Research's 2021 blog post on data cascades says they often originate early in an ML system's lifecycle, at data definition and collection. Source
  • For high-risk AI systems, Article 10 of the EU AI Act requires training, validation and testing data to be subject to data governance and management practices that include data-preparation operations such as annotation, labelling, cleaning, updating, enrichment and aggregation. Source
  • Reddit's Form 10-K for fiscal 2025 states that, to date, substantially all of the contract value associated with its licensing revenue is derived from two partners. 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