Can You Trust AI Tools for Auditors?

Overview

Most audit AI is judged on a single question: can it answer? We think that’s the wrong test. The harder and more valuable skill is knowing which disclosure questions an AI can answer reliably, and which it should hand to a person.

At the beginning of 2026 we analyzed 1.69 million disclosure-checklist questions across German GAAP, Austrian GAAP, IFRS and ESRS, and measured how often a human reviewer accepted each answer without correction. On the questions our system chooses to answer automatically, reviewers accepted its answers more often than auditor-prepared ones: 98.17% against 92.68%, roughly four times fewer corrections. It earns that by not answering everything. When it isn’t confident, it produces a proposal and defers to the auditor. We call this Calibrated Answering, and it is what ai/checklist is built around.

What you’ll learn

• Why “how often is the AI right?” is the wrong test, and why “does the AI know when it’s right?” is the one that matters.
• What calibration means in practice: auto-answer versus proposal, and why the distinction tells reviewers where to spend their judgment.
• Benchmark data across German GAAP, Austrian GAAP, IFRS and ESRS: 98.17% reviewer acceptance on AI auto-answers against 92.68% for human-prepared ones.
• What Calibrated Answering changes on the engagement, and why it matters more under the CSRD and ESRS disclosure wave.

What’s inside

1. Introduction: the wrong number most vendors quote
2. The wrong test: quality over quantity
3. What calibration actually means
4. What we measured: 1.69 million questions
5. “But aren’t those just the easy questions?”
6. What it changes on the engagement
7. What we’re not claiming, and where this leaves things

Who it’s for

Audit partners, engagement managers and innovation leads weighing where AI fits into disclosure review, and what a trustworthy, calibrated deployment looks like across German GAAP, Austrian GAAP, IFRS and ESRS, with the auditor reviewing, approving and signing throughout.

What we are not claiming

The paper is explicit about the limits of its own numbers, and so is this page. 98.17% is reviewer acceptance, not correctness against ground truth. The factor of four is fewer corrections, not a precision multiple. We are not claiming the AI is a better auditor than a person. The claim is narrower: on the questions it chooses to answer automatically, its answers are accepted more often. Nothing reaches the file without an auditor’s review, approval and signature. If you want the reasoning behind that restraint, it is in the FAQ and in the paper itself.