Synthetic data can lower barriers to testing and research, but its market value is constrained by measurable usefulness, privacy risk and the work required to prove both. This Finndy Research model separates those dimensions and presents transparent fictional cost scenarios; it is not a market-price survey or a claim of regulatory certification.
Start with the decision that synthetic data must support
Synthetic datasets do not have one universal economic value. A table that reproduces common transaction patterns may be useful for testing a dashboard while remaining unsuitable for evaluating rare fraud events. Before quoting a price, define the buyer’s task, target population, sampling horizon, permitted downstream uses and acceptance thresholds. A test-only license and a production model-evaluation license should not be treated as interchangeable products. Record the real-data baseline used for comparison, but avoid disclosing private records in public examples. The seller should state which transformations and source permissions were used, which variables were excluded, and which statistical properties are not promised. A compact acceptance sheet might require agreement within five percentage points for selected aggregate rates, while separately testing minority segments and rare events. That five-point tolerance is a hypothetical contract term, not an industry rule. Price follows the evidence that a dataset meets the agreed purpose, not merely its row count.
Separate production, verification and privacy evidence
Build three cost ledgers. Production includes source preparation, generator development, compute, reproducibility scripts and versioned exports. Verification includes schema checks, downstream benchmark execution, rare-group diagnostics, manual review and bug remediation. Privacy evidence includes threat-model review, membership-inference testing where appropriate, documentation of residual risk, and any independent assessment promised in the contract. These activities have different owners and renewal cycles, so hiding them inside a single per-million-row fee obscures actual margins. The NIST SDNist tool, described in the official software record last updated February 1, 2022, illustrates why synthetic-data assessment should cover both utility and privacy metrics. Tool output is evidence for particular tests, not a certification that a dataset is anonymous. For every release, retain the model version, test configuration, benchmark dataset version and the approval record. Buyers then have a way to distinguish an updated dataset from an undocumented regenerated file.
Calculate a price floor with transparent assumptions
Consider a purely hypothetical one-off enterprise order, not observed market prices. Scenario A is a developer test set with $500 of generation work, $250 of utility checks, $300 of privacy review and $50 of delivery. Direct cost is $1,100; adding a 20% support-and-rework reserve produces a $1,320 indicative cost floor. Scenario B supports internal research with stronger cohort tests: $750 generation, $500 utility checks, $600 privacy review and $150 delivery. Direct cost is $2,000, and the same reserve produces a $2,400 floor. Scenario C involves an extensive production evaluation: $1,000 generation, $900 utility tests, $900 privacy review and $200 delivery. Direct cost is $3,000, producing a $3,600 floor with the same reserve. The arithmetic is reproducible: floor = (generation + utility testing + privacy evidence + delivery) × 1.20. The 20% reserve is a planning assumption, not a claim about typical supplier margins. Final price also depends on license scope, buyer-specific value, competition, support obligations and bargaining. Never quote these illustrative amounts as a public benchmark.
Avoid turning synthetic into a privacy guarantee
Synthetic does not automatically mean anonymous. A generator can reproduce unusual source records or leak identifying patterns; the risk depends on the method, the training data and the attacker model. NIST’s 2021 discussion of differentially private synthetic data distinguishes mathematically bounded protections from synthetic techniques that offer no formal guarantee. Suppliers must not label a dataset differentially private unless the actual generation mechanism and its parameters support that assertion. The European Data Protection Board’s Guidelines 02/2026 on Anonymisation, published for consultation on July 8, 2026, provide additional regulatory context; a consultation document should not be misrepresented as a final compliance certificate. Describe the claimed protection precisely, retain evaluation evidence, and specify whether a buyer may combine the file with other datasets. If privacy requirements force larger utility losses, show that tradeoff in the offer rather than charging for an unrealistic perfect-privacy promise.
Measure utility on buyer tasks, not visual similarity
A synthetic chart can look convincing while failing the buyer’s workflow. Use task-specific tests such as model ranking, calibration, rule coverage, missing-value handling, subgroup error and stress cases. State the reference cohort, test-set leakage controls, holdout period and acceptance rule before generating candidate outputs. For example, an internal dashboard prototype may need stable totals and category frequencies, whereas a fraud-scoring sandbox may need credible tail behavior and class imbalance. A buyer should be able to rerun the tests independently without receiving unnecessary identifiable source data. Distinguish factual metrics from hypothetical illustration: a demonstration with 85% utility on a defined task is not an 85% overall accuracy guarantee. Record failed metrics and rejected versions. If improving one cohort increases error in another, report the distribution instead of collapsing it into a single marketing score. Evaluation effort is a separately measurable input into pricing and a strong reason for higher-priced service tiers.
Contract for rechecks, refreshes and exit
Turn the calculations into a commercial acceptance plan. Specify number of included regeneration rounds, how schema changes are billed, who pays for retesting after buyer pipeline changes, and the response window for a documented defect. A recurring subscription may be appropriate for quarterly refreshed data, but it must specify which validation work is repeated and which tests are sampled. Separate an optional privacy-audit fee from the generation fee so neither party pays twice for the same evidence. When a dataset fails pre-agreed acceptance criteria, use a defined repair, replacement or refund process rather than hiding the failure behind an aggregate quality score. Provide an export manifest, reproducibility hash, version history and clear restrictions on retention and redistribution. Review gross margin after actual rework, not just at initial quote. This approach gives the seller defensible costs and the buyer a verifiable utility-and-risk rationale instead of a speculative price per row.
Sources
- NIST, SDNist: Synthetic Data Report Tool (software record last updated February 1, 2022)
- NIST, Differentially Private Synthetic Data (2021)
- European Data Protection Board, Guidelines 02/2026 on Anonymisation, public consultation opened July 8, 2026
