When AI Prepares a Regulatory Filing, Who Attests That It Is True?

After AI Enters the Workflow · Season Three: “When AI Starts Acting for the Organisation” · Article Eight

A business must lodge a report with a regulator by the end of the month. Previously, several departments supplied data, compliance staff reconciled it item by item and an authorised officer read the final document before making a declaration. AI can now extract from financial systems, customer records, incident logs and earlier filings. It completes forms, explains material changes and can even log into a portal to submit.

Two weeks of work become two hours. Only one button remains: “Confirm that the contents are true and accurate, and submit.” The responsible officer faces hundreds of finished pages without knowing which figures came directly from source systems, which were transformed by AI, which explanations were generated by a model or whether time remains to reconstruct the checks.

What does the click attest? That AI ran successfully? That the document looks plausible? Or, in the authority of the officer’s role, that each material statement has an adequate basis?

A filing can be prepared and technically transmitted automatically. Attesting that it is true is a separate institutional function. It should not be compressed into a final ceremonial click merely because earlier production became faster.

A filing contains four layers of evidence

A regulatory document is not simply text. It converts institutional facts into statements an external authority may rely on. At least four layers should remain distinct:

Source records
→ transformation and classification
→ facts and explanations in the filing
→ declaration or certification by an authorised person

Source records include transactions, events, accounts, contracts and operational logs. They can themselves be incomplete or wrong.

Transformation and classification determine period, category and regulatory definition and handle duplication, omission and exception. Many material errors occur here rather than during prose generation.

Filing statements turn data into outward claims: “No material incident occurred” or “the change was mainly caused by this factor”. Such statements mix calculation and judgement.

Declaration or certification is made by a person with legal, professional or organisational authority and represents that the filing meets a required standard.

AI may participate in the first three layers without gaining the identity or responsibility of the fourth. A model has no officeholding authority and does not face internal discipline, regulatory action or professional consequence for a false declaration. The organisation must identify who attests, the scope of that attestation and the checks that provide a reasonable basis.

Automated extraction does not make the evidence self-proving

When AI retrieves a number from a system, “source verified” can appear automatic. The business must still ask whether the system is authoritative, covers the complete period, uses consistent unit, currency and time zone, includes later corrections and avoids duplicate representation of one event.

Generative AI may also classify unstructured material—for example, deciding an incident category from complaint text or whether minutes contain a disclosure matter. That is not copying. It is a classification decision. The filing should expose the rule, confidence boundary and human review rather than present the result as a raw fact.

The Australian Taxation Office’s business record-keeping guidance explains that records need to explain transactions and tax affairs and satisfy requirements concerning accessibility and intelligibility. Australian Taxation Office, “Record keeping for business” Particular filing duties vary, but the common design principle is traceability from a final figure to supporting records. A plausible AI explanation cannot replace that chain.

Review material assertions and changes, not every finished page

Requiring the officer to recompute hundreds of pages defeats automation. Showing only a summary and asking for a click cannot support meaningful attestation. A better workflow presents an “assertion package” for each material claim.

It includes:

• reporting requirement or definition;
• source system and cut-off time;
• transformation, exclusion and estimation method;
• change from prior period or expectation;
• exact AI involvement;
• automatic and human checks completed; and
• unresolved differences and their owner.

The officer concentrates on change, exception, material estimate and management judgement. Stable standard fields can be supported by automated reconciliation and sampling. This is risk-based certification, not delegation of responsibility to AI.

The interface should also permit rejection of all or part of the filing and return an issue to a specific data owner. A system offering only “confirm all” at the deadline turns compliance pressure into rubber-stamping.

Do not merge data ownership, preparation and attestation into one vague role

A reliable process has at least three forms of ownership.

The data owner maintains definitions, completeness and correction in the business system. Finance, operations, people and security functions own their domains.

The filing owner understands the regulatory requirement, maps internal information to fields and manages explanation and timing.

The attestor has formal authority and makes an outward declaration on the basis of required checks.

One person may hold more than one role, but the system should preserve the distinction. Otherwise, after failure each participant says they owned only the tool, source data or final submission, and no one owns the conversion from internal fact to external claim.

The Australian Government’s responsible AI policy requires covered agencies to appoint accountable officials, maintain use-case inventories and undertake impact and risk management. Australian Government, “Policy for the responsible use of AI in government” Commercial obligations differ, but explicit roles and use-case ownership are equally valuable in AI-assisted filing. Tool ownership does not replace reporting accountability.

Automated checks should find contradictions, not only formatting errors

AI can perform systematic pre-filing tests that go beyond empty fields:

• whether totals reconcile with detail and ledger;
• whether different reports classify the same event consistently;
• whether “zero incidents” conflicts with logs or complaints;
• whether explanation cites missing or obsolete policy;
• whether a large change has recorded support; and
• whether an estimate uses an approved method and is labelled.

Reverse tests are also useful. Sample a material number in the filing and trace it to records; sample events from source systems and confirm that they entered the report. The first tests accuracy and the second completeness. Looking backward only from the finished report misses events never included.

The NIST AI Risk Management Framework emphasises measurement, monitoring and documentation of limits, with governance roles able to respond to risk and feedback. NIST AI Resource Center, “AI RMF Core” In reporting, evaluation should test coverage, classification and traceability from records to required fields, not merely whether prose is correct.

Submission credentials should remain separate from preparation

Ability to prepare a filing does not imply permission to lodge it. Submission uses an institutional credential, triggers deadlines and creates an official external record. If AI prepares, clears warnings and submits, it can solidify its own mistake without independent control.

A safer design has AI create a submission package; a policy layer validates approvals, hash, version and time; and an authorised person or independent submission service transmits. The organisation then retains the regulator’s receipt, final file and exact version so later editing cannot overwrite what was submitted.

The UK National Cyber Security Centre’s secure deployment guidance stresses access control, logging, monitoring and incident management. NCSC, “Secure deployment” Regulatory portal credentials should not be persistently exposed to a model that can ingest external text or generate arbitrary tool calls.

A deadline does not reduce the standard of attestation

Automation is often justified by avoidance of late filing. Time pressure also makes people more likely to trust complete, fluent output. The system should show unresolved matters continuously before the deadline rather than produce an apparently finished document at the end.

When evidence is missing, the organisation needs pre-agreed lawful options: label an estimate, seek an extension, correct after lodgement or pause. AI should never invent a number for a mandatory field or convert unknown into “none” merely to complete the form.

Evaluation should record timeliness, corrections, material differences, open exceptions, review time and regulatory feedback. Measuring hours saved alone rewards fast submission rather than demonstrable truth.

An error requires finding every claim that used the same evidence

If an omitted source record is found after submission, correcting one figure is insufficient. Data lineage should identify every filing, period and regulator that used the same record, transformation or classification rule.

Correction involves stopping the faulty source, preserving the original filing, having an authorised person determine notice and amendment, updating connected records and verifying receipt of the correct version. Whether formal correction is required and how it occurs depend on the reporting regime and should not be decided by a general AI.

The incident should also change attestation. If the officer reviewed only aggregates while the error came from classification, future review needs classification samples. If data ownership was unclear, record responsibility must be repaired before training the model to explain differences better.

Conclusion: AI can prepare the evidence, but cannot become an ownerless attestor

Regulatory filing turns internal fact into an outward, accountable statement. AI can improve collection, reconciliation, classification, explanation and formatting without acquiring the standing to certify the result.

The final principle is:

Every material filing statement should be traceable to source records, transformation rules and completed checks. Final attestation must come from a person with the required authority, visibility of material change and unresolved issues, and power to refuse submission.

Mature automated reporting does not help an officer click faster. It lets the officer know precisely what is being attested, which evidence supports it and what remains uncertain. AI can shorten document production. It cannot shorten the distance between an institution and responsibility for truth.

Primary sources and further reading

• Australian Taxation Office: Record keeping for business
• Australian Government: Policy for the responsible use of AI in government
• NIST AI Resource Center: AI RMF Core
• UK National Cyber Security Centre: Secure deployment

Continue reading: Explore the After AI Enters the Workflow series.