Who sets the accountability bar in AI-assisted advisory work, and where is that bar heading?
Adrian E. Urgena · Founder & CEO, Mezisa
AI is taking on more of the analytical work behind advisory services: the diligence findings, valuations, quality-of-earnings analyses and restructuring plans on which clients, lenders and investors base their decisions. More and more of that work arrives with the same reassurance: a human was in the loop. It appears in vendor marketing, in engagement letters and in the methodology pages of advisory firms large and small. It is meant to answer the question every one of them eventually asks: can I rely on this?
Claiming “human in the loop” does not answer the question. It describes a workflow. It says someone was present. It does not say who, whether they were qualified for the question, whether they saw enough to judge it, or whether anything they did could have changed the answer. Those are the elements of accountability, and the parties who decide what counts are now asking about each one.
Presence is not judgment
My last piece argued that clients will keep paying a third party for qualified judgment applied to analysis: judgment that sharpens it, and then stands behind the result as fit for action. The difficulty is that judgment is invisible from the outside, and presence is easy to claim.
A reviewer who approves a thousand agent outputs a day is in the loop. So is a partner who skims a summary of work produced by a system nobody on the engagement fully understands. Neither is accountable in any sense a lender, an investor or a board would recognize.
The plainest statement of the problem comes from the supply side. In its registration statement filed in October, Anthropic names human review of AI agent output as a control that may fail, because the pace and volume of agent actions can exceed reviewers’ capacity to evaluate them. It also warns that it may face claims alleging it is liable for harm caused by its agents’ actions, even where the customer configured, deployed or supervised the agent.¹ The company selling the capability has disclosed to public investors that nominal oversight is a risk, not a safeguard.
Who sets the bar
AI vendors present human review as the safeguard. Anthropic’s usage policy requires human oversight of high-stakes automated decisions, and its product for financial advisors builds the advisor’s review into each workflow “to help firms meet their compliance requirements.”¹ Whether that review is enough, however, is not up to them. Rather, the bar is being set by the bodies that regulate professional work and the insurers that cover it, and they are moving in one direction.
Audit and tax first. Regulation reaches audit and tax first, because both already require a named professional to sign the work: the auditor signs the opinion, and the preparer signs the return. Publishing what it describes as the first guidance from any audit regulator on generative and agentic AI, the UK’s Financial Reporting Council stated that the human auditor is always accountable and that its AI guidance does not change that position.² Singapore’s audit regulator followed in October, stating that “AI can assist, but the auditor owns the outcome” and that this responsibility “cannot be delegated to AI.”³ Releasing its AI practice aid for financial statement auditors that same month, the AICPA’s chief auditor stated that “ultimately, the auditor remains responsible for the audit and the audit opinion.”⁴ In US tax practice, the IRS Office of Professional Responsibility has told practitioners that AI cannot substitute for their own review.⁵
Advisory next. The same position is now reaching advisory work, through the bodies that set its standards. In July, staff of the International Ethics Standards Board for Accountants issued a publication addressed to all professional accountants, not only auditors: practitioners remain responsible for their judgments and decisions, and must apply appropriate human oversight to what technology produces.⁶ Australia’s standard-setter for accountants is drafting AI guidance for valuation and forensic services, and its staff report that those services continue to rely on human judgment.⁷ Restructuring practitioners point the same way. When R3, the UK restructuring and insolvency body, surveyed its members, the obstacles they named were not about what AI can do. They want certainty that their professional indemnity insurance covers AI-assisted work, and clear guidance on what adequate human oversight looks like.⁸
Insurers. Carriers are moving too, and at least one does not treat human review as enough to keep AI-assisted work covered. The US insurer W. R. Berkley has introduced an absolute AI exclusion for its errors-and-omissions, D&O and fiduciary liability products, removing cover for claims “based upon, arising out of, or attributable to” any “actual or alleged use, deployment, or development of Artificial Intelligence.”⁹ Errors-and-omissions insurance is the professional liability coverage advisory firms commonly carry against claims that their advice or services were negligent, erroneous or incomplete. The exclusion turns on whether a claim arises from the use of AI, not on whether a person reviewed the output. Closer to the profession, the AICPA’s member insurance program makes the same point in its risk guidance: a risk control director at CNA, the program’s underwriter, warns that relying on AI output without checking it “may lead to errors or omissions and potential professional liability claims.”¹⁰
None of these bodies accepts presence as proof. Each is asking for something more specific: a qualified person, a review that was real, and a record that shows it.
Where the bar is heading
The direction is toward demonstrable judgment: judgment that could have changed the answer. Published requirements and guidance already map to three questions any buyer can ask today.
Was the person qualified for this question? The AICPA’s practice aid outlines engagement partners’ responsibilities to understand the AI tools they use and to ensure the firm has approved them.⁴
Could they have changed the answer? The IRS expects practitioners to review AI-generated content thoroughly before it reaches a client, verifying facts, citations and calculations.⁵ ACRA puts the test to auditors directly: “Does your independent reasoning corroborate the AI-generated conclusion, or are you relying solely on the AI to reach that conclusion?”³ Review that cannot keep pace with the work, as the Anthropic S-1 concedes it may not, does not meet that test.¹
Can they show it? The IRS recommends documenting AI use and the verification steps taken.⁵ ACRA’s guidance says audit documentation should show “how AI-assisted outputs were reviewed” and “how professional judgement was applied.”³ Stakeholders told Australia’s standard-setter that users of valuation and forensic work should be informed when and how AI has been used.⁷
Today’s requirements mostly record that review took place. The next step is evidence that the review mattered, and ACRA has already pointed there: its guidance suggests audit documentation may also include “significant revisions made following human review, and the rationale for accepting or rejecting AI-generated recommendations.”³ Judgment that could not have changed the answer is not judgment. It is attendance.
Certifying the machine
A second route to trust is growing alongside professional review: certifying the AI system itself. In August KPMG became the first Big Four firm with an AIUC-1 certified AI capability.¹¹ In October the AICPA and the Artificial Intelligence Underwriting Company launched a joint engagement that pairs a SOC 2 examination with live testing of AI agents.¹² The route has institutional backing, and it will grow.
Its sponsors are careful about what it proves. A KPMG principal told CFO Dive that certification is “evidence of rigor” but does not “eliminate or provide immunity from risks,” and that maintaining confidence will require “continued monitoring, testing and human accountability.”¹³ The AICPA’s version is attested by a licensed CPA. Certification tests a system’s controls and behavior, not whether a particular output is fit for a particular decision. It builds an assurance layer beneath the professional, not a replacement for one. The risk lies in how it is read: a buyer who treats “certified AI” as a stand-in for a professional who stands behind the output has mistaken a better machine for a judgment about reliance.
What would challenge this
If a lender accepts AI-produced diligence with no qualified professional standing behind it, if an E&O carrier gives credit for a certified system in place of evidence of human review, or if a regulator accepts automated oversight of professional deliverables, this argument would be in serious doubt. None of those is visible today.
For firms, “human in the loop” is about to stop working as an answer. For buyers, better questions are already available: who, qualified how, and what did they change? Accountability was never about who was in the room. It is about whose judgment could have changed the answer, and who will stand behind the answer they gave.
About this piece
Drafted with AI assistance from Mezisa’s market research. Every factual claim is sourced below. Reviewed, revised and adopted by Adrian E. Urgena, who stands behind it.
Adrian E. Urgena is founder and CEO of Mezisa Systems. He has over 25 years of operator and business advisory experience across more than 150 engagements and $30B in M&A deal value. He holds a BA in Economics from Stanford and an MBA from the University of Denver.
Sources
Anthropic, PBC, Form S-1 registration statement (preliminary prospectus), filed October 26, 2026: pp. 47, 50–52 (risk factors on agent oversight and liability), p. 203 (Claude for Financial Advisors), pp. 231–232 (usage policy). https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm
Financial Reporting Council, “Innovative new guidance supports audit firm adoption of emerging AI technologies” (March 30, 2026). https://www.frc.org.uk/news-and-events/news/2026/03/innovative-new-guidance-supports-audit-firm-adoption-of-emerging-ai-technologies/
Accounting and Corporate Regulatory Authority (Singapore), “Audit Practice Guidance No. 1 of 2026: Using Artificial Intelligence Responsibly” (issued October 2, 2026), pp. 4–6. https://isomer-user-content.by.gov.sg/277/4bde1c7a-331f-4967-82ec-7c4c31d186cf/Audit_Practice_Guidance_No1_of_2026_Using_Artificial_Intelligence_Responsibly.pdf
AICPA & CIMA, “AICPA releases guidance to help auditors address the use of AI in financial reporting and auditing” (October 13, 2026). https://www.aicpa-cima.com/news/article/aicpa-releases-guidance-to-help-auditors-address-the-use-of-ai-in-financial-reporting-and-auditing. Also reported by Journal of Accountancy, “AI and the audit: AICPA publishes practice aid to assist auditors” (October 13, 2026), https://www.journalofaccountancy.com/news/2026/oct/ai-and-the-audit-aicpa-publishes-practice-aid-to-assist-auditors.html
IRS Office of Professional Responsibility, Alert 2026-19, “Introductory Guidelines for Responsible AI Use in Federal Tax Practice” (June 24, 2026), as summarized by Mondaq. https://www.mondaq.com/unitedstates/new-technology/1823662/irs-office-of-professional-responsibility-issues-introductory-guidelines-on-artificial-intelligence-use-in-federal-tax-practice
IESBA, “New IESBA Staff Publication Highlights Ethical Considerations for Accountants Using Emerging Technologies” (July 15, 2026). https://www.ethicsboard.org/news-events/2026-07/new-iesba-staff-publication-highlights-ethical-considerations-accountants-using-emerging. Summary: ICAS, https://icas.com/news-insights-events/news/ethics/iesba-issues-guidance-on-the-ethical-implications-of-emerging-technologies/
APESB, Agenda Item 9, “Update on AI Project: Forensic Accounting and Valuation Scoping” (staff paper, May 2026). https://apesb.org.au/wp-content/uploads/2026/05/Agenda_Item_9_Update_on_AI_Project_Forensic-Accounting-and-Valuation_Scoping.pdf
R3, “R3 report: Restructuring and insolvency professionals want clearer guidance and safeguards to support AI use” (September 2, 2026). https://www.r3.org.uk/news/r3-report-restructuring-and-insolvency-professionals-want-clearer-guidance-and-safeguards-to-support-ai-use/
Hunton Andrews Kurth, Insurance Recovery Blog, “The Continued Proliferation of AI Exclusions” (May 28, 2025). https://www.hunton.com/hunton-insurance-recovery-blog/the-continued-proliferation-of-ai-exclusions
Sarah Beckett Ference, CPA, risk control director, CNA, “Generative AI and risks to CPA firms,” AICPA Member Insurance Programs (2023). https://www.cpai.com/Education-Resources/My-Firm/Data-Security-Risk-Management/Generative-AI-and-risks-to-CPA-firms
KPMG LLP, “KPMG LLP becomes first Big Four firm with AIUC-1 certified AI capability” (August 27, 2026). https://kpmg.com/us/en/media/news/kpmg-llp-becomes-first-big-four-firm-with-aiuc-1-certified-ai-capability.html
CPA Practice Advisor, “New AICPA joint offering provides full review of business risk related to AI systems and agents” (October 19, 2026). https://cpapracticeadvisor.com/2026/10/19/new-aicpa-joint-offering-provides-full-review-of-business-risk-related-to-ai-systems-and-agents/191955
Alexei Alexis, “KPMG gets agentic tool certified for security, reliability as AI risks mount,” CFO Dive (September 1, 2026). https://www.cfodive.com/news/kpmg-agentic-tool-certified-security-reliability-ai-risks-mount/829260/
Start with a conversation.
Whether you’re weighing a growth, performance or exit decision, or want to compare notes on where advisory is going next, we welcome the conversation.
