Our PCAOB comment letter on AI-assisted audit evidence

Acceptable AI-assisted audit evidence should be defined by three properties, not by an approved list of tools: every conclusion traces back to the specific evidence it rests on, the inputs and model versions are documented well enough for a reviewer to re-perform the procedure, and a documented human review shows who remains accountable for the conclusion. On 7 August 2026, DNL Deep Neuron Lab GmbH filed that position with the Public Company Accounting Oversight Board in response to its Request for Public Comment on PCAOB Standard Setting (Release No. 2026-005).

The letter argues for staff guidance rather than a full standard-setting cycle, because firms are choosing their audit methods this season and a multi-year cycle will not reach them in time. Its central case is that standards written without AI in mind currently make the weaker audit the safer one to defend: a firm that tests 100 percent of journal entries carries inspection risk a firm that samples does not. We have a commercial interest in the answer and the letter says so in its second paragraph. dnl’s own tooling is documented on the ai/checklist and ai/numbers product pages, and our earlier read of the UK regulator’s position is in What the FRC’s AI guidance requires, and where firms fall short.

The full text as filed follows. Footnote markers in the original are rendered here as the Sources list at the end.

The letter as filed

August 7, 2026

Office of the Secretary
Public Company Accounting Oversight Board
1666 K Street, NW
Washington, DC 20006-2803
Via email to comments@pcaobus.org

Re: Request for Public Comment on PCAOB Standard Setting (PCAOB Release No. 2026-005)

Dear Office of the Secretary:

Thank you for the opportunity to comment and for opening the Board’s standard-setting agenda to public input. It is the right moment to ask the question we address here: what should acceptable AI-assisted audit evidence look like? This letter responds principally to Questions 3 and 8–10, on the use of AI and the choice between staff guidance and standard setting, and to Questions 17–19, on alignment with international auditing standards.

We are Deep Neuron Lab GmbH, trading as dnl, a Berlin-based technology company building AI-assisted tools for financial audit. Our tools are used by audit firms across the United Kingdom, continental Europe, Australia, and Canada, from independent practices to Big Four member firms. We offer this comment as observers of how firms elsewhere are resolving the question now before the Board. We have a commercial interest in how the Board approaches this question, and we state that plainly. What follows is written from what we observe in audit practice rather than from a position of neutrality we do not hold.

Our view: the most constructive action the Board can take at this stage is to issue staff guidance. AI adoption in audit is outpacing multi-year standard-setting cycles, which is why staff guidance is the right first instrument, informing the eventual standard rather than replacing it.

When the safer choice is a weaker audit (Questions 8–10)

Consider a firm that uses AI to test 100 percent of a company’s journal entries rather than taking the traditional manual sampling approach. One outcome is that regulators recognize the fuller coverage as an improvement. The other is that, absent clear standards, the firm cannot predict what an inspector will ask about the model, the data, and the method. In turn, the firm retreats to manual sampling because it is easier to defend from a compliance standpoint. The rational choice becomes the weaker audit.

The scenario is not ours. Last year, then-Board member Christina Ho set it out in her remarks “Building Tomorrow’s Audit: Four Strategies for a Tech-Forward Future” at the World Continuous Auditing and Reporting Symposium,[1] and her closing question was the right one: does moving from full-population testing back to a sampling approach advance investor protection?

The firm that adopts the stronger method carries the risk; the firm that stays with the status-quo manual method is safe. When the safe choice is also the lower-quality choice, standards written without AI in mind work against the very thing the Board exists to protect. This is not hypothetical or far off. Firms face this choice now, in live audit seasons, without an AI-specific benchmark for what good AI-assisted work should look like.

The Board is not starting from nothing here. In 2024 it amended AS 1105 and AS 2301 to address technology-assisted analysis, and those amendments do real work, but they were not written for AI specifically, which is the gap that remains.[2] Three examples (Question 10):

• AS 2315 (audit sampling). Full-population testing strains its sampling concepts, which assume selection from a population rather than examination of all of it.
• AS 1215 (audit documentation). It does not yet say what documentation of a model’s inputs, versions, and outputs would make an AI-assisted procedure re-performable on inspection.
• AS 2810 (evaluating audit results). It requires the auditor to evaluate whether the financial statements contain the information essential for a fair presentation. It does not say what would support reliance on a tool that checks presentation and disclosures against the full population of applicable requirements, rather than a manually maintained checklist.

The Board has recognized this gap: its research agenda includes a data-and-technology project whose stated objective is to assess whether guidance or standards changes are warranted as auditors adopt new tools. Our comment is directed at that project, and our recommendation is that it produce staff guidance sooner, rather than as the output of a full standard-setting cycle; no timeline has been published, and adoption is accelerating.

Our proposal: guidance built on properties, not products (Questions 3, 8–10)

To date, the Board has wisely avoided endorsing a single model, tool, or vendor, and we are not asking it to start. But the guidance gap is real, and action is needed. Guidance can instead describe the properties that make AI-assisted evidence trustworthy, leaving firms free to meet them however they choose. That description should be technology-neutral and lean on three principles:

Traceability. Every AI-assisted conclusion can be traced back to the specific evidence on which it rests.

Reproducibility. The inputs, model versions, configurations, and outputs are documented in enough detail that a reviewer can re-perform and evaluate the procedure, because re-performance is what an inspection depends on.

Human accountability. A documented record shows a human auditor reviewed the work and remains responsible for the conclusion. The tool proposes; the auditor decides and signs.

To make this concrete, consider disclosure completeness. A tool that checks a set of financial statements against the full population of applicable disclosure requirements, rather than against a manually maintained checklist, does fuller work than the method it replaces. Under these three properties, the auditor relying on it would show which requirement each flagged item traces to, would retain the requirement set, inputs, and outputs so a reviewer can re-perform the check, and would document the auditor’s own resolution of every flagged item. Nothing in that record asks an inspector to understand the model’s internals; everything in it can be inspected.

An auditor who can demonstrate their work in these three ways should be able to rely on an AI-assisted procedure with reasonable confidence that it will withstand inspection, whatever technology produced it. We offer these three properties as a starting framework the Board could refine within firms’ existing obligations under AS 1105, AS 1215, and QC 1000. This keeps professional judgment and accountability with the human auditor while giving firms the certainty they need to adopt methods that expand audit coverage rather than retreat to narrower ones. In other words, the technology changes; the auditor’s responsibility does not. We would note that this concerns the auditor’s use of AI to obtain evidence, which is a distinct question from an issuer’s use of AI in preparing the financial statements. Both merit the Board’s attention.

Regulators are converging (Questions 17–19)

The Board asked how it might achieve greater alignment with international auditing standards, and its draft 2026-2030 Strategic Plan proposes fostering that alignment as an objective. The answer is unusually clear right now. On August 5, the IAASB published proposed revisions to ISA 330, ISA 500, and ISA 520 addressing auditors’ use of technology, taking a principles-based rather than prescriptive approach; the comment period runs to December.[3] The UK’s Financial Reporting Council has gone furthest: its Generative and Agentic AI Guidance (March 2026), following its AI in Audit paper (June 2025), is explicit that responsibility for AI deployment and output quality rests unchanged with the human auditor.[4] We raise this not to hold up another jurisdiction as a model the Board should adopt, but because it shows guidance of this kind can be written quickly without displacing the auditor’s accountability. Germany’s Chamber of Public Accountants (WPK) maintains a question-and-answer framework on the use of AI in audit practice, updated most recently in May 2026.[5] The international direction is being set as we write.

Aligning with this direction, rather than developing a separate U.S. approach, would spare the firms the Board oversees from conflicting requirements on the same AI-assisted procedure. We are not proposing a single global rulebook. We are proposing that regulators moving into the same territory learn from one another rather than diverge, which is both more achievable and, ultimately, in the interest of the U.S. investors the Board protects.

Our recommendation (Question 8)

We respectfully urge the Board to issue staff guidance on acceptable AI-assisted audit evidence now, anchored to traceability, reproducibility, and human accountability, rather than awaiting a full standard-setting cycle that may run for years. Firms are choosing their methods today. Staff guidance would remove the disincentive to adopt stronger methods, give investors the benefit of fuller testing, and keep the human auditor firmly accountable while the longer process runs its course. The cost of waiting is not neutral; it is measured in audit quality forgone, season after season.

We would welcome the opportunity to share aggregated, anonymized observations on adoption patterns at the practitioner level, without reference to any firm or engagement. We appreciate the Board’s attention and stand ready to engage in any further dialogue.

Respectfully submitted,

Andreas Schindler
Chief Executive Officer
Deep Neuron Lab GmbH

Sources

1. Christina Ho, “Building Tomorrow’s Audit: Four Strategies for a Tech-Forward Future,” remarks at the World Continuous Auditing and Reporting Symposium (WCARS), Rutgers University, November 7, 2025. The journal-entry scenario appears in the section “Strategy #2: Using AI in Audit.” Ms. Ho served as a PCAOB Board member until January 2026.
2. AS 1105, Audit Evidence, and AS 2301, The Auditor’s Responses to the Risks of Material Misstatement. The Board adopted amendments to both standards addressing technology-assisted analysis in June 2024 (effective for audits of financial statements for fiscal years beginning on or after December 15, 2025).
3. International Auditing and Assurance Standards Board (IAASB), “Audit Evidence and Risk Response — ISA 330, ISA 500, ISA 520,” proposed revisions published for public comment on 5 August 2026 (comment period through 15 December 2026).
4. Financial Reporting Council, “AI in Audit” guidance (first guidance published June 2025; Generative and Agentic AI guidance published March 2026), described by the FRC as the first such guidance from any audit regulator globally.
5. Wirtschaftsprüferkammer (WPK) [German Chamber of Public Accountants], “Fragen und Antworten zum Einsatz von künstlicher Intelligenz in der WP-Praxis” [Questions and Answers on the Use of Artificial Intelligence in Audit Practice], as of 21 May 2026, announced 15 June 2026. The English translation lags the German original and remains at the 26 May 2025 version.