On Thursday, the PCAOB posted its 2025 inspection reports for the six largest US audit firms, and the numbers are unlike anything the modern inspection regime has produced. EY, whose inspectors found deficiencies in 28 percent of reviewed audits a year ago and 37 percent the year before that, came in below 5 percent. Deloitte, below 5 percent as well, from 14. PwC landed at 9 percent, KPMG at 12.5. Even the tier below moved: Grant Thornton fell from 48 to 33 percent, BDO from 60 to roughly 34. The figures come from the inspection reports as tallied by Accounting Today and Thomson Reuters, with the underlying data on the PCAOB’s own dashboard.

A quick decoder for readers who do not live in this world. In these inspections, a Part I.A deficiency is the serious kind: the inspection team concluded the firm signed an opinion without having obtained sufficient appropriate evidence to support it. It does not mean the financial statements were wrong, and it is a long way from fraud; it means the file, as inspected, did not yet justify the signature on it. The PCAOB reviewed 64 audits at each Big Four firm. So EY’s move from 28 percent to below 5 means roughly eighteen deficient audits in one cycle and three in the next. That is not a rounding improvement. That is a different firm, at least as the inspection lens sees it.

The question that matters for everyone who buys, signs, or relies on an audit is why. The firms have an answer ready, and it is the one you would guess.

EY’s Americas assurance vice chair, Joe Link, called the results historic and said they “validate the investments we’ve made in technology, data analytics, streamlined methodologies and our people.” The firm points to a $1 billion program that, per an Accounting Today feature the same week, includes its Helix audit data platform running in over 150 countries, AI automation that the firm says touches 130,000 of its assurance professionals, and its analytics leader’s estimate that AI accelerates assurance work by 125 to 150 percent. Deloitte credited “a human-led, AI-powered approach that enhances our professionals’ experience and judgment through technology-enabled precision and scalability.”

Give the story its due, because the mechanism is genuinely plausible. Most Part I.A deficiencies are, at bottom, evidence gaps: a test not performed, a population not covered, a contradiction not chased. The strongest argument the pro-AI camp has ever made about audit quality, the one a PCAOB board member used to make before she left the board, is that machines can test entire populations instead of samples. If that is what Helix-style tooling actually does across 64 inspected audits, fewer evidence gaps is exactly the result you would predict. Even the cautious wing of the regulator conceded the point last winter: the then-acting chair, in the same speech that warned about eroding skepticism, acknowledged that AI can already strengthen risk assessment and evidence gathering by analyzing entire populations. On the firms’ telling, 2025 is the year the promise showed up in the regulator’s own numbers.

It might be true. Before you repeat it in a pitch meeting or a board deck, three other things moved at the same time, and none of the coverage we read this week mentioned any of them.

First, the trajectory predates the AI deployments. Big Four deficiency rates fell from 26 percent to 20 percent in the 2024 cycle, an improvement the PCAOB’s then-chair Erica Williams claimed as the fruit of years of pressure: “We challenged the audit profession to do better for America’s investors,” she said in March 2025, when the trend was already visible. The remediation programs, coaching, and methodology overhauls that follow ugly inspection years were launched in 2022 and 2023, when EY’s rate peaked at 46 percent, before generative AI touched most audit files. EY’s billion-dollar program itself was announced years before this cycle’s results, which cuts both ways: long investments take time to land, and so does ordinary remediation. Some of what the firms now attribute to technology is the slower compounding of getting yelled at by a regulator for four consecutive years.

Second, the lens changed hands mid-measurement. These 2025 inspections were fielded and written up across the exact period we covered in our regulator edition: Williams resigned in July 2025, an acting chair bridged six months, and a rebuilt board arrived in January 2026 under the banner of sensible, efficient oversight of auditors. The new chair, Demetrios Logothetis, described the current approach this way in the coverage of these very reports: “When we inspect at the quality control level, we are not only looking at whether an audit was performed appropriately. We are evaluating whether the system… is functioning in a way that consistently produces high-quality audits.” That is a legitimate philosophy of inspection. It is also a different philosophy than deficiency-hunting audit by audit, and no one outside the process can fully separate how much of a 28-to-5 move reflects better audits versus a changed way of looking at them. We are not alleging softer grading. We are saying that when the thermometer and the temperature change in the same year, a careful reader holds both possibilities.

Third, the denominators are small and the mix moves, and you do not have to take our word for it. The PCAOB prints the warning on its own data page: the audits selected do “not constitute a representative sample of the firm’s total population of issuer audits,” selection mixes “risk-based and random methods” that focus on a different mix of audits from year to year, and “inspection results are not necessarily comparable over time or among firms.” Sixty-four audits per Big Four firm; at those sizes, three audits separate a headline of “below 5 percent” from one of “nearly 10.” The direction across all six firms is too consistent to dismiss, but the precision the percentages imply is not there, and the regulator says so itself.

Our read: something real improved. Six firms moving the same direction, two cycles in a row, is not noise, and it would be strange if a billion dollars of tooling and several years of remediation produced nothing. But the causal story being marketed, AI did this, is precisely the version that cannot be verified from outside, and it happens to be the version that sells software, justifies the spend, and flatters every party involved, including a regulator that would like its new approach to coincide with good news. When every actor in a system benefits from the same explanation, that explanation deserves extra scrutiny, not less.

The test is next year. If the rates hold near 5 percent through another cycle, under a settled inspection regime, with the same firms disclosing how the tooling actually changed the work, the AI story earns the credit it is claiming this week. If the rates drift back toward the teens, 2025 will look like what skeptics of one-year miracles always suspect: a confluence.

If you sit on an audit committee or buy an audit, expect these numbers in every proposal you see for the next two years. The follow-up questions are simple and revealing: which specific procedures on our engagement are now performed differently because of the technology, what does the human review of machine-selected evidence look like, and what did your firm’s rate look like across the last four cycles, not one.

If you are a practitioner at a large firm, notice what the winning explanation implies about your job. Both EY’s and Deloitte’s statements put the humans at the center of the story, experienced professionals supervising powerful tooling. That is the version of this profession the inspection numbers now appear to reward, and it is the skill set to build deliberately rather than by accident.

If you rely on audited financial statements, the honest summary is that the trend is good news for the reliability of what you read, whatever its cause turns out to be. Just keep the decoder in mind when the numbers get quoted at you: fewer evidence deficiencies is not a guarantee that every set of statements is right, and the causal story is where the marketing lives.

If you run or work at a smaller firm, the uncomfortable part of Thursday’s release is the gap. Below 5 percent at the top, 33 and 34 percent at the next tier down. Whatever combination of money, tooling, and attention produced the Big Four’s numbers, it is capital-intensive, and the inspection scoreboard is now publicly widening the quality distance between those who can afford it and those who cannot. That gap will show up in marketing decks long before it shows up in standards.

The best available reading of Thursday’s reports is that audit quality genuinely improved, that technology contributed, and that nobody, including the firms, can honestly decompose how much came from the machines versus the remediation grind versus the regulator’s new lens. Hold the good news and the epistemic humility at the same time. And notice the shape of the explanation everyone converged on: human-led, AI-powered, the professional’s judgment amplified rather than replaced. Even in the profession’s best week of inspection results in memory, nobody dared claim the machine did it alone, because the entire system, from the engagement letter to the inspection report, still runs on a person taking responsibility for the file. AI can do the work. It cannot sign the work. On the evidence of this week, the firms that internalized both halves of that sentence are the ones the inspectors could barely fault.

— Footnote

Footnote is an independent publication. It is not professional accounting, tax, or legal advice. Deficiency figures are from the PCAOB’s 2025 inspection reports as compiled by the outlets linked above, with underlying data on the PCAOB’s public dashboard; firm statements and performance claims are attributed to the companies and are not findings we audited. Our discussion of possible explanations is analysis, explicitly including possibilities we cannot verify from outside the inspection process, and we allege no impropriety by any party. Details are current as of August 18, 2026.