Related posts
See all
Key Insights:
- Firms using AI that touches client data and signed opinions face independence and skepticism risks beyond the scope of generic enterprise AI policies.
- No binding PCAOB, AICPA, or SEC standard specifically governs AI use in audit engagements yet, but inspectors are already scrutinizing AI-intensive audits.
- Effective AI governance at an audit firm requires layering complementary frameworks rather than relying on any single standard, since none alone addresses program, system, and assurance-level needs together.
A senior manager pastes a client's trial balance into a chat tool to draft a risk narrative. A staff auditor uses an agent to review a vendor contract that sits inside a confidentiality agreement. A partner signs an opinion supported by workpapers an agent helped assemble. None of those moments existed in the firm's quality manual three years ago, and most firms have not written down who is accountable for them now.
That gap is what governance has to close. AI use at an audit and advisory firm sits inside engagement evidence, workpapers, review workflows, signed opinions, and client confidentiality at the same time, which makes governance an immediate practice issue rather than a future policy exercise. This article covers why governance at an audit firm looks different from generic enterprise AI policy, the three frameworks worth layering, and the operating model that holds the program together as AI use expands.
Why AI governance looks different at an audit firm
Most AI governance guidance published in the last two years was written for general enterprise AI use: customer-facing chatbots, demand forecasting, HR screening tools. That guidance centers on bias, privacy, and operational reliability. Those risks matter for audit and advisory firms too, but the profession carries an additional layer: confidentiality obligations, workpaper integrity, independence, and professional skepticism.
AI tools used in engagement workflows touch client data covered by confidentiality agreements and produce or inform workpapers behind signed opinions. All of that sits inside a regulatory environment where independence and skepticism are foundational to the profession's credibility.
The PCAOB, AICPA, and SEC have not issued a binding AI-specific audit standard, but the issue is already on regulators' radar. The PCAOB has raised it publicly, inspectors are monitoring emerging tools in audits, and in December 2025 the SEC signaled it is reconsidering the independence framework. Firms that build governance now, using the best available frameworks, will face less disruption when binding standards arrive.
The frameworks that matter and how to layer them
No single AI framework was built for audit and advisory firms, but three together cover what you need: NIST AI RMF for system-level risk, ISO/IEC 42001 for program-level management, and the IIA's AI Auditing Framework for assurance-specific guidance. Layering them gives you a structure that fits engagement work rather than internal-only AI use.
NIST AI RMF: The system-level risk lens
The NIST AI RMF is the framework most firms reach for first because it gives leaders a vocabulary for risks the audit standards don't yet name: hallucination, training data drift, prompt injection, sensitive information leakage. Its four functions (Govern, Map, Measure, Manage) separate firmwide accountability from the controls that have to exist on every individual AI system.
That separation is what makes it useful in practice. A firm can have a polished AI policy and still fail an inspector's question about how a specific tool was tested before it touched a client engagement. NIST forces both conversations: who owns the program, and what evidence exists that each system is behaving as intended in the workflows where it runs.
ISO/IEC 42001: The program-level management system
ISO/IEC 42001 is the first international standard that lets a firm get its AI program certified the same way ISO 27001 lets it get its information security program certified. Where NIST gives you a risk vocabulary, 42001 gives you the management system: defined roles, documented controls, internal audits, continuous improvement, and an external assessor who signs off.
For audit and advisory firms, that certifiability matters in two ways. It produces evidence a client, regulator, or insurer can verify without taking the firm's word for it, and it forces the internal discipline (control owners, evidence cycles, recurring reviews) that AI governance needs to survive partner turnover and tool churn. Fieldguide is among the first audit and advisory platforms to hold ISO 42001.
IIA AI auditing framework: The assurance-specific lens
The IIA's updated AI auditing framework is the one written by auditors for auditors. Where NIST and ISO describe how an organization should govern AI, the IIA framework describes how an independent assurance function should evaluate that governance, mapped to the Three Lines Model that audit leaders already use.
That distinction is what makes it work in two directions for a firm. Internally, it gives the internal audit or quality function a tested structure for reviewing the firm's own AI use, with criteria a partner can defend in a peer review. Externally, it is the closest thing the profession has to a playbook for the emerging client service line: SOC-style engagements where CPAs evaluate a client's AI systems against NIST AI RMF or ISO 42001 as criteria.
Governing two categories of AI, not one
Your firm needs separate governance for a chat tool that a practitioner queries on demand and an agent that executes multi-step engagement work independently. Human-orchestrated AI and agent-executed AI raise different questions about authorization, tracing, and review.
Human-orchestrated AI, where the practitioner decides when to trigger an action and reviews the output immediately, needs acceptable-use policies, data handling rules, and quality checks on outputs. It resembles governing any analyst tool with a judgment overlay.
Agent-executed AI, where an agent drives the workflow and a practitioner reviews the results afterward, introduces new governance territory. Decision rights shift. You are defining what the agent is authorized to do, what configuration controls apply, how outputs are traced, and what the review standard looks like.
If your firm writes a single AI policy covering both categories, it will often over-constrain the simple tools or under-govern the consequential ones. Separating governance by category, with shared principles but distinct control sets, keeps the program workable as AI use expands.
A practical operating model
Frameworks tell a firm what good governance looks like; the operating model decides whether it actually happens. Most firms are still building that model in real time, often while the AI governing council is still figuring out what the real risks are. Three structural decisions tend to determine whether the program holds up as AI use spreads across the firm.
Cross-functional ownership
Governance that sits only with IT will write rules that engagement teams quietly route around. The people who understand independence obligations, client confidentiality, and engagement economics have to be on the committee alongside IT and risk, which is why recent audit leadership research keeps surfacing practice leadership as a missing seat. A working AI council usually pairs a practice partner, a CIO or head of technology, a quality or risk leader, and someone close to the engagement workflow itself.
Decision rights that scale
The hardest part of expanding governance is not writing the policy; it is deciding who reviews what and when, without routing every new AI use case to the same small committee. The pattern that holds up at scale, supported by a recent responsible AI survey, is to build accountability into the design and deployment workflow itself and split responsibilities by risk tier. High-risk systems that touch client opinions or evidence get full council review; lower-risk tools (an internal summarizer, a marketing draft) move through a lighter approval path, so the committee's attention stays on what actually carries professional risk.
Engagement-level champions
A governance program lives or dies at the engagement level, because that is where the AI tools actually get used. KPMG's dual training approach is a useful model: teams are trained both to evaluate clients' AI processes and to use AI tools responsibly inside their own engagements. In practice, the people carrying that load are managers and senior associates, who need to know what documentation the firm expects, what falls inside policy, and when a situation should be escalated rather than handled at the engagement level.
Human oversight, independence, and professional skepticism
Even with the right frameworks and operating model in place, AI governance ultimately rests on whether reviewers actually challenge the outputs in front of them. The PCAOB has flagged automation bias (the tendency to accept a confident-looking machine output rather than test it) as a direct risk to audit quality and professional skepticism. The danger is not that AI produces bad work; it is that polished AI output looks finished, which makes the reviewer less likely to push on it.
General skepticism training does not solve this on its own. Reviewers tend to catch more when they apply a specific cognitive protocol, often called a counterarguing mindset, where they actively look for reasons the AI output could be wrong before accepting it. For a firm, that means updating review checklists and training to include structured counterargument steps for any workpaper an AI tool helped produce, not just relying on a general instruction to "review carefully."
Independence is the other piece that does not get easier with AI in the workflow. The same independence rules apply to the AI tools the firm uses internally and the AI assurance services the firm sells to clients, and the ISQM-1 and QC-1000 quality management standards already require firms to assess the risks of new tools and document the governance that mitigates them. New AI tools fall squarely inside that obligation.
Match governance to AI maturity
Governance needs to be built for where the firm actually is, and designed to grow with where it is going. Most firms are still at early stages of AI governance maturity, often without a formal GenAI framework in place at all. The goal at every stage is the same: enough structure to manage the AI use happening today, with enough flexibility to expand as the firm adds more.
For firms early in the journey, governance usually starts with the basics: an inventory of every AI tool in use (including third-party tools embedded in other software), an acceptable-use policy, and clear thresholds for what has to be disclosed to clients or documented in workpapers. ISACA's AI governance guidance groups these foundations under principles like digital trust and oversight.
As AI use deepens, governance expands to cover the harder questions: who approves agent configurations, what the review standard looks like for agent-executed work, how the firm monitors performance over time, and how it delivers advisory and assurance services around AI to clients. Firms that design their program with those later stages in mind avoid rebuilding from scratch 18 months later, when AI has spread well beyond what a basic acceptable-use policy can cover.
Fieldguide's AI Maturity Framework maps this progression in the context of audit and advisory work specifically, connecting each stage to the governance controls and operational capabilities that make the next level attainable.
Build AI governance on a platform designed for it
Most AI tools were not built for the professional responsibility that comes with audit and advisory work, which is why governance ends up bolted on after the fact. Fieldguide is the industry's only end-to-end AI-native platform, purpose-built for audit and advisory, with the Agent Workforce, methodology depth, and audit-grade rigor firms need to operate this way. Its security foundation includes ISO 42001 certification and SOC 2 Type 2 attestation, giving firms verifiable evidence that AI governance is built into the platform itself. Half the top 100 firms use Fieldguide, including members of the Big Four, and every AI-supported output still passes through practitioner review and approval. Request a demo to see how Fieldguide supports AI controls across every stage of firm AI adoption.
Amanda Waldmann
Increasing trust with AI for audit and advisory firms.
