Skip to main content

Answering a regulator's letter: responding with documents, not assurances, anchored to real enforcement patterns

The short answer

Documents, not assurances

The single rule of this topic. A regulator acts on records it can verify, not sentences that assert compliance. The DoNotPay finding was a documentary gap: the claim existed, the test did not (FTC, 2024 to 2025). Answer every request with a file, not a phrase.

What you will be able to do

  • Evaluate a regulator's letter and separate what it is actually asking for (specific documents, specific facts, by a specific date) from what a nervous organization wants to send instead (reassurance, narrative, and promises of future work).
  • Judge whether a proposed response answers the request with evidence or merely asserts compliance, using the "documents, not assurances" test that the DoNotPay and wider Operation AI Comply pattern makes concrete.
  • Map a competent authority's request to the specific artifacts in your dossier that answer it: the systems inventory, the data provenance file, the incident and post-incident records, the evaluation report, and the conformity file.
  • Apply the EU AI Act's cooperation duties correctly: Article 21 (a provider's duty to give a competent authority the information, documentation, and system logs it reasonably requests) and Article 26(12) (a deployer's duty to cooperate with authorities), and know which one is yours for a given system (see Topic 5.1).
  • Assess what a "reasoned request" from a market surveillance authority under Article 74 can lawfully reach, including documentation and the training, validation, and testing datasets, and where the limits sit.
  • Construct a regulator response that leads with a short, accurate cover letter and an evidence index, discloses honestly, corrects the record where a past claim overreached, and never invents a fact or a document that does not exist.
  • Decide what to hold back and why: what is genuinely out of scope, what is protected, and why the answer to "should we just send everything" is almost never yes and almost never a flat no.
  • Anticipate how the response you send becomes a permanent record that a later adversary will attack, and write it so it survives that attack (see Topic 11.1).

The ideas, one by one

The letter is a scoped request, not a verdict

It names who is asking, under what authority, for what specific documents, by what date. Answer that request, exactly and completely, and resist the urge to answer the accusation you imagine instead.

Your role decides your duty

For a high-risk system under the EU AI Act, Article 21 governs providers (information, documentation, and Article 12 logs on reasoned request) and Article 26(12) governs deployers (cooperate with authorities). Know your role per system before the letter arrives (see Topic 5.1).

Reach is broad but bounded

A market surveillance authority under Article 74 can, on a reasoned request, reach documentation and training, validation, and testing datasets, and in defined circumstances source code. "Reasoned" is a real limit: answer it fully, and narrow an untethered request through counsel.

Over-disclosure is a failure, not safety

Cooperation means answering what was reasonably asked, accurately and completely, not dumping everything. Over-answering waives protections, exposes third-party and trade-secret material, and buries the helpful evidence in noise.

Correct, do not defend the indefensible

If a public claim outruns your evidence, state what the evidence supports, fix the overstated claim, and attach proof of the fix. Documented correction resolves items that stubborn defense turns into headlines.

Never improve a document after the letter arrives

Editing, backdating, or manufacturing evidence converts a survivable gap into fraud and can breach preservation duties. Send exhibits as they existed. An honest gap is survivable; a doctored record is not.

The response is a forever record

Whatever you send will be reread by an adversary, a hostile board, or an examiner against your other artifacts, hunting for contradiction (see Topic 11.1). Write every sentence so it survives that reread a year later.

The dossier is the difference

If you built the artifacts across Modules 0 to 5, answering is assembly. If you did not, answering is manufacture under deadline, which is where organizations panic, over-disclose, or fabricate. The whole program's build-as-you-go discipline exists for this moment.

This is judgment, not filing

Deciding which document answers which question, what to correct, what to withhold, and where the reasonable-scope line sits is trained evaluation on real stakes. That judgment, exercised on evidence you built, is what makes you more capable than someone who only studied the law.

Read the letter twice: for the ask and for the trigger

The literal questions define the scope of your duty; the trigger between the lines (a complaint, a headline, a marketing phrase quoted back at you) tells you what the regulator is really testing, which is almost always the gap between a public claim and its evidence. When a letter quotes your own advertising, that quoted phrase is the true subject of the inquiry, whatever the numbered questions say.

The rule crosses regulators and sectors

The same "documents, not assurances" logic that convicted DoNotPay at the FTC also drove the SEC's first AI-washing settlements, where advisers paid penalties for claiming AI capabilities they did not have (SEC, 2024). Different regulator, different sector, identical standard: a claim of capability is the exposure, and the absence of evidence for it is the violation. Expect the question wherever the letter comes from.

Consistency across parallel inquiries is a survival skill

One system can draw letters from several authorities at once, and each can eventually read what you told the others. Tell the same truth, the same way, with the same exhibits, to every one of them. A story tuned per audience is the easiest contradiction for a later adversary to find, because you already put it in writing and signed it.

Listen to it

The same lesson, talked through by two hosts as episode 40 of the podcast.

Real cases

These are real, documented enforcement patterns. Each shows the same underlying truth: regulators test the gap between what an organization claimed and what it can prove, and the organizations that fare best answer with evidence.

Example 1: DoNotPay and the missing test (FTC, 2024 to 2025). This is the module's anchor and the clearest instance of the rule. DoNotPay marketed a "robot lawyer" and made sweeping claims about replacing human legal work. The FTC's complaint centered not on whether AI can ever help with legal tasks, but on a documentary gap: the company had not tested whether its tool performed at the level of a human lawyer and had not retained attorneys to check its law-related output (FTC, DoNotPay, 2024 to 2025). The February 2025 order imposed 193,000 US dollars in monetary relief, required notice to 2021 to 2023 subscribers, and barred the unsupported "performs like a real lawyer" claim absent sufficient evidence (FTC, February 2025). The teaching point: the claim was the exposure; the absence of a test document was the violation.

Example 2: Rytr and the tool that could be misused (FTC, 2024, and its 2025 reversal). Also part of Operation AI Comply, Rytr sold an AI "Testimonial and Review" writing service that could generate detailed reviews unrelated to any real experience (FTC, Rytr materials, 2024). In December 2024 the FTC approved a final consent order, by a divided 3 to 2 vote, barring Rytr from offering review-generation services; two commissioners dissented on the ground that a tool with lawful uses should not be treated as inherently deceptive (FTC, Rytr, December 2024). This example carries a live currency lesson: on 22 December 2025 the FTC reopened and set aside the Rytr order, citing the current administration's deregulatory AI Action Plan (FTC, "FTC Reopens and Sets Aside Rytr Final Order," December 2025). The enforcement pattern of Operation AI Comply is real and instructive; a specific outcome can later be undone by a change in political posture. Teach the durable rule (documents, not assurances) and mark the shifting enforcement climate as emerging, not fixed (see Topic 5.1 for the US federal posture).

Example 3: The DeepSeek information request (Italy's Garante, 2025). When the Italian data-protection authority (the Garante) opened its DeepSeek inquiry, it began exactly as this topic describes: with a formal request for information asking five specific things (what data was collected, its sources, the purposes and legal basis, where it was stored, and how users were told), with a twenty-day deadline. The provider answered with a jurisdictional assurance ("the law does not reach us") rather than the documents asked for, and the Garante found the responses insufficient and ordered processing to stop (Garante, decisions of January 2025). This case is owned in depth by Topic 5.1 for the scope question (see Topic 5.1); here it illustrates the response failure: answering a document request with a legal assurance is not answering it.

Example 4: Clearview AI and the evidence of practice (multiple EU authorities, 2021 to 2024). European data-protection authorities in Italy, France, Greece, the United Kingdom, and the Netherlands issued penalties against Clearview AI over its facial-recognition database. Across those proceedings, what the authorities acted on was documented practice (the scale of scraping, the absence of a lawful basis, the failure to respond adequately to access requests), established through the company's own materials and public record rather than through the company's characterizations of itself (various national DPA decisions, 2021 to 2024). The pattern again: regulators build on evidence of what was actually done, so the organization that keeps honest records of what it did is the one that can respond to them meaningfully. The deep treatment of a facial-recognition conformity failure is owned by Topic 5.6 (see Topic 5.6).

Example 5: The financial-sector model examination (ongoing supervisory practice). Prudential and conduct regulators in banking and insurance across the EU, the UK, and the US routinely examine firms' use of models, including AI and machine-learning models, through supervisory reviews that request model documentation, validation reports, and governance records. Firms with mature model-risk-management practice (a documented inventory, independent validation on file, and clear ownership) answer these examinations as an assembly exercise. Firms without it scramble. The lesson transfers directly: the AI Act's document duties formalize, for AI specifically, a discipline that regulated finance has practiced for years. The organizations that already keep the file find the AI Act's Article 21 request familiar.

Example 6: The SEC "AI-washing" advisers and the capability that was never there (SEC, 2024). On 18 March 2024 the US Securities and Exchange Commission (the SEC, the American markets regulator) announced its first enforcement actions against investment advisers for false or misleading statements about their use of AI, settling separate cases with Delphia (USA) Inc. and Global Predictions, Inc. (SEC, press release 2024-36, 18 March 2024). Delphia had told investors, from 2019 to 2023, that it "put[s] collective data to work to make our artificial intelligence smarter so it can predict which companies and trends are about to make it big and invest in them before everyone else"; Global Predictions had claimed in 2023 to be the "first regulated AI financial advisor" offering "expert AI-driven forecasts" (SEC, 2024). The SEC's finding was the now-familiar shape: the firms did not in fact have the AI and machine-learning capabilities they advertised. The firms settled and paid 400,000 US dollars in total civil penalties (Delphia 225,000; Global Predictions 175,000), charged under the Investment Advisers Act's antifraud provisions (Sections 206(2) and 206(4)), the Marketing Rule, and the Compliance Rule (SEC, 2024). Two teaching points carry forward. First, this is the same "documents, not assurances" rule enforced by a different regulator in a different sector, on the same day-one logic as DoNotPay: the claim of a capability is the exposure; the absence of the capability (and of any record substantiating the claim) is the violation. Second, the Compliance Rule charge is the quiet lesson for a governance professional: the SEC faulted the firms not only for the false claim but for failing to have written policies and procedures to prevent it. A functioning claims-substantiation process is itself the kind of document a regulator looks for; its absence is its own finding.

Where people go wrong

  • "We should reassure them that we take this seriously." Reassurance is an assurance, and assurances are what regulators discount. The DoNotPay pattern is precise: confidence without a test document is the violation, not the defense. Replace every sentence of reassurance with a document that proves the point, or delete it.
  • "Send everything, so they cannot say we hid anything." Over-disclosure is its own failure. A reasoned request has a scope; answering far beyond it can waive protections, expose third-party and trade-secret material, invite new lines of inquiry, and bury the helpful evidence in noise. Cooperation means answering what was reasonably asked, completely and accurately, not dumping the company's hard drive.
  • "Let marketing or communications draft the reply." The instinct to hand a regulator a polished narrative is the instinct to send an assurance. Communications is informed; it does not own the substantive answer. Legal owns strategy and privilege; the governance owner owns the map from question to evidence; the technical owners produce the logs and tests.
  • "We can promise the testing program we are building now." A promise of future work is not evidence of past compliance. Regulators judge the state of the world when the conduct happened. If the document did not exist for the period in question, saying you are building one now does not close the gap; only honest disclosure of the gap does.
  • "If a claim is weak, defend it hard." Stubborn defense of an unsupportable claim is how a small inquiry becomes a headline finding. Prompt, documented correction (state what the evidence supports, fix the overstated claim, attach proof of the fix) is treated far more favorably and often resolves the item.
  • "We can tidy up the evaluation report before we send it." Editing, backdating, or improving a document after an inquiry begins converts a compliance gap into a fraud problem and can trigger obligations to preserve, not alter, records. Send documents as they existed. A real gap is survivable; a doctored record is not.
  • "Privilege means we can refuse to answer." Legal protection shapes how you respond, not whether you must. A lawful, reasoned request still binds you; protected material may be handled through confidential treatment (the AI Act's Article 78 supports confidentiality for information given to authorities) or appropriate redaction of out-of-scope third-party data, on legal advice, not withheld wholesale.
  • "The regulator can demand anything, so resistance is pointless." A market surveillance authority's reach under Article 74 is broad but purpose-bound and tied to a reasoned request. You answer the reasoned request fully; you may, through counsel, ask an untethered or wildly overbroad request to be narrowed. Precision is not obstruction.
  • "Cooperation is optional goodwill." Under the EU AI Act it is a legal duty with defined content: Article 21 for providers (information, documentation, and Article 12 logs on reasoned request) and Article 26(12) for deployers (cooperate with authorities), on top of the deployer's own record-keeping duties. Which one is yours depends on your role for that system (see Topic 5.1).
  • "Once we answer, it is over." Whatever you send is a permanent record. A later adversary, a hostile board, or an examiner will read it against your other artifacts hunting for contradictions (see Topic 11.1). Write every response so it survives being reread by an enemy a year later.
  • "An informal inquiry is just a friendly chat, so we can speak freely." The first contact from a regulator is often informal (a soft letter, a request for a call) before any compulsory instrument issues, and the reflex is to relax. Do not. A loose verbal assurance in the informal phase can become the exact overstatement a later formal request tests, and everything you say can shape the formal proceeding. The fix: treat the first informal contact with the same evidentiary discipline as a formal demand. Preserve records, know your role, speak only in facts you can prove; the instrument's formality changes your room to negotiate scope, not the standard of honesty.
  • "We can shade the story a little differently for each regulator." A single AI system can draw parallel letters from several authorities (data protection on the data, consumer protection or market surveillance on the claims, a sector regulator on the use), and it is tempting to tune each response to its audience. That tuning is the easiest thing in the world for a later adversary to exploit, because the contradiction is already in writing and signed by you. The fix: one truth, told the same way, with the same exhibits, to every authority. Consistency across parallel responses is not a nicety; it is what survives cross-checking.
  • "If the request is vague, we should guess the scope ourselves." Some letters ask for "all documents relating to" a system, which could mean five files or five thousand. Guessing wide invites over-disclosure; guessing narrow risks an under-answer that breaches the duty. The fix: seek prompt, documented clarification through counsel, proposing a reasonable reading of the scope in writing and asking the authority to confirm it. Early clarification is good faith, and whichever way the authority answers, you now have the scope in writing to answer against.
  • "We can answer without ever pinning down our exact role for this system." Drafting a response before confirming whether you are the provider or the deployer for the specific system is building on sand: the role fixes the duty, and the wrong role produces the wrong exhibits and either an under-answer or an over-answer (see Topic 5.1). The fix: resolve the role for this exact system first, from your applicability memo, before a single exhibit is chosen. If you rebranded or substantially modified a system you bought, note that this can turn a deployer into a provider and re-check the role rather than assuming the old label.

Questions people ask

What is regulator's letter (information request)?
A formal, authority-backed request from a regulatory body for specific documents or facts about an organization's conduct, by a specific deadline. It is a scoped request, not a verdict or a fine. In the US it often takes the form of a civil investigative demand; in the EU AI Act context, a reasoned request from a market surveillance authority.
What is assurance?
A statement that asserts compliance or quality is true ("we tested it thoroughly," "we take this seriously") without a verifiable record behind it. Regulators discount assurances. The opposite of a document. More on Assurance
What is document (evidence)?
A record that lets a regulator independently verify a claim: a test result, an evaluation report, a log, a dated change record, a signed review. The currency regulators actually act on.
What is Operation AI Comply?
The US Federal Trade Commission's September 2024 coordinated law-enforcement sweep bringing five cases against deceptive AI claims and schemes, including the DoNotPay "robot lawyer" action. The module's real anchor for the enforcement pattern that rewards documents over assurances (FTC, 25 September 2024).
What is civil investigative demand (CID)?
A compulsory pre-litigation request for documents and answers that the FTC and many US state attorneys general can issue to investigate suspected violations. The US analogue of a market surveillance authority's reasoned request.

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →