I’ve been watching the AI-in-audit story develop for a while now. Lots of headlines, lots of hype, not much regulatory clarity. That changed on March 30, 2026.
The UK’s Financial Reporting Council published what it calls — and I think is accurately described as — the world’s first formal guidance from any audit regulator on generative and agentic AI. Not a discussion paper. Not a request for comment. Actual guidance, with a defined risk taxonomy and clear expectations for firms.
The headline message is simple: you can’t blame the AI. Accountability stays with the audit partner. But the document goes much deeper than that, and that’s what I want to walk through here.
Why This Guidance Is Different
For years, AI in audit meant rules-based automation — think Alteryx, Excel macros, Python scripts. Explicit instructions, deterministic outputs, traceable failures. The human wrote the logic. The machine followed it. The risk profile was manageable.
Generative and agentic AI are fundamentally different. Ask the same prompt a hundred times and you get a hundred different answers. These systems are non-deterministic, harder to trace, and in the case of agentic AI, capable of pursuing multi-step goals with real autonomy — without a human directing every step. The failure modes don’t look like anything we’ve dealt with in traditional audit automation.
The FRC published this guidance proactively — before quality failures emerge at scale, not in response to them. That’s worth noting.
The Three Risk Categories
The FRC organizes AI risk to audit quality into three categories. The important thing to understand upfront: Categories 2 and 3 can produce audit failures even when the AI output itself is perfectly correct.
① Risk of Deficient Output
The AI produces a bad output — wrong, incomplete, or misleading. This breaks down into four sub-sources: the LLM component itself performing poorly (hallucinations, omissions, distortions, faulty reasoning, inconsistencies), non-LLM components like data connectors or OCR tools failing, errors compounding as they pass between components, and the auditor providing a bad prompt or wrong input to begin with.
② Risk of Misuse of Output
The AI output is technically fine. The auditor applies it incorrectly. This is a human failure enabled by over-reliance on the tool. An accurate summary of board minutes used to draw a conclusion the underlying minutes don’t support. An AI-generated risk flag treated as a definitive finding without professional judgment. A tool used outside the scope it was designed for — maybe it doesn’t have access to all the files it should, so it gives you a perfectly reasonable answer based on incomplete context.
③ Risk of Non-Compliant Methodology
The firm’s approved approach for using AI in audit doesn’t meet auditing standards — even when the AI output is good and the auditor uses it correctly. The failure is upstream. The firm designed or approved an AI-assisted procedure that, as written, doesn’t satisfy the relevant standard. This is invisible at the engagement level. It lives in how the firm built its methodology.
The Five LLM Deficiency Types
Within Category 1, the FRC names five specific ways an LLM’s output can be deficient. None of these should be completely new, but having them named and defined matters for how you review AI output in practice.
- Hallucinations — The model generates plausible-sounding text that is simply made up. It’s been trained to produce an answer, and sometimes it will produce one even when no real answer exists. The most famous examples are lawyers getting caught citing fabricated court cases.
- Omissions — Information that should be in the output isn’t. Models are getting better at finding the needle in the haystack, but they’re not perfect. In trying to optimize token usage, they sometimes deprioritize content that’s highly relevant to us but that the model deems less important.
- Distortions — The facts are present but their significance is altered. Understated, overstated, or reframed. This tends to happen when the model has less context than it should and makes a decision to focus on the wrong area.
- Faulty Reasoning — Unsupported or illogical conclusions. Studies on sycophantic AI behavior show that models often agree with the user even when the user is wrong, which can contribute directly to this.
- Inconsistencies — Outputs that contradict themselves or prior outputs. If you keep telling the AI it’s wrong, it will often flip its answer — even if the original answer was correct.
These five deficiencies trace back to four root causes: lack of semantic awareness (the model generates statistically plausible text without truly understanding meaning), dependence on training data (biased or inaccurate training data produces biased or inaccurate outputs), finite model capacity (attention allocated unevenly, some content dropped), and finite context window (only so much fits in active memory at once).
The Three Agentic-Specific Risks
This is where it gets more interesting for where audit is heading. Agentic AI — systems that pursue goals across multiple steps with some autonomy, calling tools, integrating outputs, managing their own iteration — introduces risks that simply don’t exist in a standard chat session.
Think of the difference between asking Claude a question in a chat window versus telling Cowork to go do something. The chat is a single prompt you review immediately. Cowork builds a plan, runs around doing things, and you see the result at the end. Very different risk profile.
Goal Drift
As an agentic system runs through a long multi-step task, its context window fills up. To keep running, it has to compress older content — and sometimes that includes the original goal or parts of the work plan. By the end of the workflow, the system may be optimizing for a subtly different objective than what was set at the start. This doesn’t happen in a single chat prompt. It’s a risk unique to long-running agentic tasks.
Amplification Risk
A small error introduced early in the pipeline — uploading the wrong file, an incorrect first step — can cascade through every subsequent component. Each step processes the slightly-wrong output and potentially makes it worse. By the time a human sees anything, what started as a minor issue has become a significant deficiency. In a single-prompt GenAI tool, each output is a fresh start. In an agentic pipeline, outputs feed other outputs.
Interface / Combination Risk
Agentic systems connect to a lot of things — different tools, data sources, external systems. Semantic mismatches (one component produces output in a format another can’t correctly interpret) or data transfer failures between components can produce deficient outputs even when every individual piece works correctly in isolation. The AI’s goal is to finish the task, so it may keep running on bad inputs rather than stopping to flag the problem.
What This Means in Practice
Knowing this taxonomy changes how you review AI output. An auditor who understands that LLMs can hallucinate, omit, and distort — and that agentic systems can experience goal drift — approaches AI-assisted work differently than one who just trusts the result.
The FRC’s ask isn’t to stop using AI. It’s to design human-in-the-loop checkpoints around where specific risks are highest, document how AI outputs were validated (not just that an AI tool was used), ensure staff understand the failure modes, and certify tools before deployment. For agentic systems especially, that means identifying which steps in a pipeline carry the most amplification or goal-drift risk — and putting human review there, not just at the final output.
Key Takeaways
- The FRC’s March 2026 guidance is the first from any global audit regulator on generative and agentic AI
- Three risk categories: deficient output, misuse of output, non-compliant methodology — Categories 2 and 3 can cause failures even when the AI is correct
- Five LLM deficiency types: hallucinations, omissions, distortions, faulty reasoning, inconsistencies
- Three agentic-specific risks: goal drift, amplification, and interface/combination risk
- Human-in-the-loop checkpoints must be designed around where specific risks are highest — not assumed to happen at the end
- Accountability for audit quality stays with the engagement partner, regardless of what the AI does
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply