Leadership and readiness
Is data startup assurance investment a toll or a cost of delivery?
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Split it before judging it. Compliance work done for each dataset, such as rights review, de-identification checks and provenance records, is a cost of delivery and should fall per accepted dataset. Entity-wide assurance is a toll: fund it when a named contract requires it, at the narrowest scope the buyer accepts. DoD put one small firm's three-year CMMC Level 2 certification cycle at $104,670, assuming the controls already existed.
Value drivers
Data startup assurance
Key takeaways
- Assurance spending buys either the right to bid (a toll) or the right to deliver (a per-dataset cost); a funding plan should show the two on separate lines.
- DoD estimated a small entity's CMMC Level 2 certification at $101,752, or $104,670 over three years, against $34,277 for the self-assessment many contracts accept, and assumed the controls were already built.
- The DFARS CMMC rule applies for three years only to contracts that contain the requirement, so the toll attaches contract by contract.
- In NIST's public baselines, the moderate set is 1.93 times the low one and the high set 2.48 times, while 58 of the low baseline's 149 controls already carry a recurring frequency.
- Track fixed assurance cost per accepted delivery and compliance cost per accepted dataset by cohort; both should fall as the business grows.
What is a data startup assurance investment actually buying?
One of two things, and an investor should make the company say which. Some assurance spending buys the right to bid: a buyer's procurement rule will not let the company compete without an attestation, an assessment or a certification. The rest buys the right to deliver: rights review, de-identification checks, provenance records and buyer-specific documentation that each dataset needs before a buyer can accept it.
The two behave differently on a funding plan. A toll is a fixed cost per legal entity and system boundary, paid on the assessor's calendar whatever the volume. A cost of delivery rises with every dataset and should fall per unit as the work becomes routine. Blending them into one compliance line hides whether spending follows accepted deliveries or simply adds overhead before the first contract that needs it.
This article prices the toll with the Department of Defense's published estimates, measures how scope and recurrence multiply it using NIST's machine-readable control catalogs, ties per-dataset work to buyers' documentation duties, and ends with an allocation tree, a milestone table, the risks and the questions to put to any data company, SourceX included.
What does a buyer-imposed toll cost? The CMMC rule's own estimates
More than the headline figure suggests, because the government's estimate covers the assessment, not building the controls. In its final CMMC program rule (Federal Register, October 15, 2024), DoD estimated a small entity's Level 2 certification assessment and affirmation at $101,752, and $104,670 over three years including two annual affirmations; for entities other than small ones, the certification assessment estimate is $112,345.
Two details in the same rule matter more to an investor than the totals. A small entity's Level 2 self-assessment was estimated at $34,277, or $37,196 over three years, and DoD says many contracts involving controlled unclassified information will require only that self-assessment. The rule also assumed no added engineering cost for Level 2, because it assumed contractors had already implemented the NIST SP 800-171 Rev 2 requirements; a company starting from nothing pays for that work on top. The rule names assessor availability and the size and complexity of the enterprise or enclave as drivers of assessment cost.
Timing comes from the companion acquisition rule. The DFARS CMMC rule, published September 10, 2025 and effective November 10, 2025, applies CMMC requirements for its first three years only to solicitations and contracts that contain a CMMC requirement; after that, they apply where the contractor's own systems will process, store or transmit the covered information. The toll attaches contract by contract, which is the plainest argument for paying it when a named contract asks and not before.
Few AI data companies sell to defense buyers, but the pattern carries over. A procurement rule names the artifact, the assessment level and the cadence; the buyer, not the vendor, decides which level a contract needs; and, as with CMMC, a cheaper self-attested route may exist for the lower tiers.
How much does scope multiply the work? Counts from NIST's public catalogs
A great deal, through two dials: the assessment level a buyer demands and the boundary the assessment covers. NIST's public baselines show the first dial. Counting the machine-readable SP 800-53 Rev 5.2.0 baselines that NIST publishes on GitHub (oscal-content commit f0de6d8, October 22, 2026), the low baseline selects 149 controls and enhancements, the moderate 287 and the high 370: the moderate set is 1.93 times the low one and the high set 2.48 times.
These baselines govern federal system authorizations, not the commercial attestations most startups seek first, whose criteria are not published in machine-readable form, so the counts are a public proxy for how scope drives effort rather than a price for any commercial report. The second dial works the same way. The assessed boundary, meaning which entity, systems, locations and data an assessment covers, is a spending decision, and a company that puts its whole stack in scope when a buyer would accept an enclave around the delivery pipeline pays to evidence controls no buyer asked about.
The nonfederal catalog is smaller. NIST's SP 800-171 Rev 3 catalog in the same repository lists 130 requirement entries, 33 of them withdrawn, leaving 97 active requirements. CMMC Level 2 is built on Rev 2, so a company preparing for it should confirm which revision its contract cites before scoping any work. The counts, with the share of each set that asks the organization to fix a frequency:
| Control set (NIST public catalog) | Controls in scope | Size against the low baseline | With an organization-set frequency |
|---|---|---|---|
| SP 800-53 Rev 5.2.0, low baseline | 149 | 1.00x | 58 (38.9%) |
| SP 800-53 Rev 5.2.0, moderate baseline | 287 | 1.93x | 75 (26.1%) |
| SP 800-53 Rev 5.2.0, high baseline | 370 | 2.48x | 81 (21.9%) |
| SP 800-171 Rev 3, withdrawn entries dropped | 97 | Not comparable | 24 (24.7%) |
Why assurance never finishes: recurrence is written into the controls
Because a large part of the work is a repeat on a schedule the organization must set and then keep. In the low baseline, 58 of the 149 selected controls (38.9 percent) define at least one parameter about frequency, such as how often a policy is reviewed, accounts are checked or an assessment runs. The moderate and high baselines add only 17 and 23 more such controls to that core, reaching 75 and 81.
That shape says something about spending plans. The recurring core arrives with the first and smallest scope; moving up a baseline mostly adds depth through enhancements (18 in the low baseline, 110 in the moderate, 182 in the high) rather than new calendars. A plan that budgets assurance as a one-time project understates the recurring cost from the first year, while a plan that buys the largest scope to be safe pays mostly for enhancements no contract has requested.
The CMMC estimates show the same pattern in money: the small-entity Level 2 certification figure rises from $101,752 to $104,670 over three years because the period includes two annual affirmations. Regulators are also changing how the repeats are evidenced. FedRAMP now publishes its 2026 consolidated rules as a machine-readable dataset of definitions, process rules and Key Security Indicators, which its repository says describe capabilities and evidence expectations. Automated evidence can cut the cost of each repetition; it does not end the repetition, and the tooling is paid for before any buyer pays.
Which compliance work is a cost of delivery?
The work one dataset needs before one buyer can accept it: confirming the supplier may license the records, checking de-identification against the agreed method, documenting provenance and answering the buyer's regulatory questions. It grows with deals, so it belongs in unit economics beside preparation and delivery, not in overhead.
Buyers' own duties push this work toward suppliers. Under the EU AI Act, providers of general-purpose AI models must make public a sufficiently detailed summary of the content used for training (Article 53(1)(d)). The Commission's template for that summary has a data-sources section that distinguishes public and non-public sources, scraped data, user data and synthetic data, and providers of models placed on the market before the obligations applied must publish summaries no later than 2 August 2027. A supplier that hands over a source description fitting that structure saves the buyer a step; one that cannot leaves the buyer to rebuild it.
The investor metric is compliance cost or hours per accepted dataset, by stage and by cohort. If it falls as checklists, templates and tooling mature, the work is becoming a repeatable delivery capability; if it climbs with each new buyer, the company is absorbing bespoke review it has not priced. No private company's figure, SourceX's included, is public, so this is a question to ask rather than a benchmark to quote.
The assurance allocation tree and milestone table
Fund each assurance line only after it passes five questions in order, then release money in tranches tied to evidence a counterparty can confirm. The tree comes first, then the Illustrative milestone table; the triggers and stop rules are editorial choices for an investor to adjust, not measured thresholds.
Work through these questions for every proposed assurance line:
- Which named buyer or contract requires it, and in what words? With no named requirement, keep a self-attested baseline and stop here.
- Which artifact does that requirement accept: a self-attestation, a third-party assessment, a certification or a government authorization? Fund the cheapest accepted artifact; the CMMC self-assessment path shows how wide the gap can be.
- What is the smallest boundary the buyer accepts: one entity, one enclave around the delivery pipeline, one data flow? Scope the work to it.
- What recurs after the first assessment, how often and at what cost? Put that recurring core in the operating budget, not the project budget.
- Which event will show the money removed transaction uncertainty: a buyer's security review closed, a contract signed that named the requirement, an accepted delivery? Name it before spending.
| Tranche (Illustrative) | Release trigger | Boundary | Evidence that closes it | Stop rule |
|---|---|---|---|---|
| 0. Baseline hygiene | Any buyer questionnaire | Policies, access and logging for the delivery pipeline | Questionnaires answered with no exception a buyer rejects | No third-party report yet |
| 1. Per-dataset controls | First signed dataset deal | Rights, de-identification and provenance checks for each dataset | Accepted delivery with no compliance rework | Pause new buyer types if hours per dataset rise two cohorts running |
| 2. Third-party assessment | A named buyer requires one in writing | The smallest boundary that buyer accepts | That buyer's security review closed | No second assessment until another buyer needs a different one |
| 3. Recurring program | Two or more contracts depend on the artifact | Same boundary, renewed on the assessor's cycle | Renewals or repeat orders that cite it | Shrink the boundary if no contract cites the wider one |
| 4. Government authorization | A contract vehicle that names the level | An enclave sized to that contract | Award under that vehicle | Never ahead of a named vehicle |
Can assurance spending outrun accepted delivery?
Yes, and that is the strongest case against funding assurance early. It adds fixed cost on a calendar; compliance hires and evidence-automation tools add more; and accepted deliveries arrive only when buyers sign. A company can finish an assessment, staff a program and automate its evidence while revenue still waits on a first contract.
The opposing view deserves its due: some buyers will not open a conversation without a third-party report, so waiting for a named requirement can lose deals that never reach the pipeline. The milestone table handles that by counting written requests. When several prospective buyers ask for the same artifact before signing, that is a named requirement and tranche 2 opens.
Investors can test the plan with two ratios over successive periods: fixed assurance cost per accepted delivery, which should fall as deliveries grow, and per-dataset compliance cost per accepted dataset, which should fall as the work standardizes. When either rises for several periods in a row, spending is ahead of the evidence.
Risks that could make the toll-or-delivery split wrong
Five developments would weaken the split this article recommends. Each is worth asking about directly, because a company that has not considered them has not tested its own spending plan.
- Buyers could move ahead of contracts. If large AI buyers begin requiring the same assessment from every data supplier, the toll becomes a precondition for any conversation, and spending before a named contract would be justified.
- Prices and requirement sets move. The CMMC figures are estimates from October 2024 that assume the controls already exist; the rule names assessor availability as a cost driver, and a change in the underlying NIST revision changes what must be met.
- The proxy has limits. NIST baselines are federal control sets; commercial attestations use other criteria, so the counts show how scope multiplies work, not what any commercial report costs.
- Documentation duties can travel by contract. The EU training-content summary obligation falls on general-purpose AI model providers, not data suppliers, but buyers can pass documentation demands down in their agreements and raise the cost of every delivery.
- Assurance is not supply. A security assessment says nothing about whether a company may license the records it sells, so an investor who treats it as evidence of rights or demand misreads both.
Assurance questions for SourceX and any data company
Put the same questions to every data company, SourceX included; good answers name contracts, boundaries and cohorts rather than badges.
For SourceX, the published process is where to start. Its process covers rights review, contracting and data preparation for each licensed dataset; the supplying company approves scope and terms and signs before any data reaches SourceX; and large deliveries stay in the supplier's own storage or travel on encrypted drives instead of being hosted by SourceX. Those facts describe a process, not what it costs, and SourceX publishes no compliance spending or per-dataset figures; it shares company and commercial metrics with investors directly. The company is raising from accredited investors under Rule 506(c), and its investor materials are available on request through the investor page. Questions worth sending:
- Which assurance artifacts has a named buyer or contract required in writing, and which were bought ahead of any requirement?
- What boundary does each assessment cover, and would a smaller one have satisfied the buyer?
- What is the recurring annual cost of each artifact, and which contracts cite it?
- What were compliance hours per accepted dataset, by cohort, over the last four quarters?
- How does keeping large deliveries in suppliers' storage change the systems an assessment has to cover?
- Which rule change or buyer policy would force a wider boundary, and what would it cost?
Methods note: control counts, cost estimates and what was not measured
The control counts are this article's own, made on 2026-10-27 with a script that reads only the JSON data files of usnistgov/oscal-content at commit f0de6d8 (October 22, 2026): the SP 800-53 Rev 5.2.0 catalog, its low, moderate and high baseline profiles, and the SP 800-171 Rev 3 catalog. Baseline membership is the list of control identifiers each profile includes, base controls and enhancements alike, each counted once. A control carries a frequency parameter when one of its own parameters contains the text frequenc in its label, usage, guidelines or choices; an enhancement does not inherit its parent's parameters. SP 800-171 entries marked withdrawn are dropped. Percentages and ratios are computed from integer counts and rounded half up once.
Dollar figures are DoD's estimates in the October 15, 2024 CMMC program rule as shown in Federal Register search excerpts; they were not recomputed or adjusted. The allocation tree and milestone table are editorial frameworks drawn up on 2026-10-27, and their triggers and stop rules are Illustrative. Not measured: what any company, SourceX included, spends on assurance; the cost of any commercial attestation or certification; whether assurance changed any buyer's decision; and how many controls a given startup already meets.
Frequently asked questions
How much does a CMMC Level 2 certification assessment cost a small company?
DoD's October 2024 final rule estimated $101,752 for a small entity's Level 2 certification assessment and affirmation, and $104,670 over three years with two annual affirmations. The estimate assumes the NIST SP 800-171 Rev 2 requirements are already implemented, so building the controls costs more.
Should an early AI data company buy a third-party security assessment before its first buyer?
Usually not. Keep a self-attested baseline, scope per-dataset controls into the first deal, and buy a third-party assessment when a named buyer requires one in writing, at the smallest boundary that buyer accepts. Several prospective buyers asking for the same artifact counts as that requirement.
What is the difference between an assurance toll and a compliance cost of delivery?
A toll is entity-wide assurance a buyer requires before it will contract, paid per boundary on a fixed calendar. A cost of delivery is the compliance work each dataset needs before acceptance, such as rights, de-identification and provenance checks, and it should fall per dataset as the work standardizes.
How should investors measure compliance efficiency at a data business?
Track fixed assurance cost per accepted delivery and compliance hours or cost per accepted dataset, by cohort, over several periods. Both should fall as the business grows; a rising trend means spending is running ahead of evidence.
Does a security assessment show a data company can license its data?
No. An assessment covers controls on a defined boundary; it says nothing about whether the company holds rights to license the records or whether buyers want them. Rights and demand need their own evidence.
Sources
- DoD's CMMC program final rule (October 15, 2024) estimates a small entity's Level 2 certification assessment at $101,752 ($104,670 over three years with two annual affirmations), a Level 2 self-assessment at $34,277, and assumes the SP 800-171 Rev 2 requirements are already implemented. Source
- The DFARS CMMC final rule (September 10, 2025, effective November 10, 2025) applies CMMC requirements for three years only to solicitations and contracts that contain them. Source
- NIST's OSCAL SP 800-53 Rev 5.2.0 catalog defines each control and enhancement with its organization-defined parameters; the counts here are this article's. Source
- NIST's OSCAL low baseline profile lists the 149 SP 800-53 Rev 5.2.0 controls and enhancements it selects. Source
- NIST's OSCAL moderate baseline profile lists the 287 SP 800-53 Rev 5.2.0 controls and enhancements it selects. Source
- NIST's OSCAL high baseline profile lists the 370 SP 800-53 Rev 5.2.0 controls and enhancements it selects. Source
- NIST's OSCAL SP 800-171 Rev 3 catalog lists 130 requirement entries, 33 of them withdrawn, each with organization-defined parameters. Source
- FedRAMP's rules repository publishes the 2026 consolidated rules as a machine-readable dataset of definitions, process rules and Key Security Indicators. Source
- The European Commission's explanatory notice and template set out the public summary of training content that general-purpose AI model providers must publish under Article 53(1)(d) of the AI Act. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.