The Accounting Podcast

A solo accountant can complete two years of bookkeeping in a few hours using Claude Cowork. But KPMG had to pull an entire AI report after 89% of its citations turned out to be fake. The AI revolution in professional services is already sorting winners from losers.

In episode 493 of The Accounting Podcast, hosts Blake Oliver and David Leary tackled a cluster of stories that paint a clear picture of how AI is restructuring professional services right now in real workflows, real paychecks, and real embarrassments for the Big Four.

There’s a growing divide between professionals who use AI carefully with human oversight (and get massive productivity gains) and those who rush to market themselves as AI experts while failing basic verification. This episode covers the Big Four’s repeated AI failures, the incredible productivity gains available to practitioners who use AI right, and the broader industry signals showing how AI is reshaping everything.

The Big Four’s “Vibe Citation” Problem Keeps Getting Worse

KPMG’s 2025 report, “Total Experience: Redefining Excellence in the Age of Agentic AI,” was supposed to showcase its AI expertise. Instead, it became the latest example of Big Four firms publishing AI-generated content they apparently never checked, also known as “vibe citations.”

GPTZero, a platform originally built to help teachers detect AI-generated text, analyzed KPMG’s report and found out of 45 citations, only five were accurate. Twenty-eight pointed to real sources but had made-up details. Twelve were too vague to verify. At least 16 were complete hallucinations. The tool rated the report 89% flawed.

The fake details weren’t subtle. KPMG claimed an Austrian utility called Verbund was using AI for real-time household energy optimization. In reality, the citation was about Verbund investing in a startup that might do this someday. They said Emirates airline had a chatbot named “Sara” that could change flights. Sara was actually a robot assistant from 2023 with no flight-change capability. The biggest gaffe was claiming East Japan Railway was using AI agents in 2019, before this type of AI even existed commercially.

UBS, NHS Greater Manchester, and Transport for London all said KPMG’s claims about their use of AI were “completely false or misleading.”

“We have to create a database and just track these because the Big Four just keeps doing it over and over again,” David said, noting similar recent incidents at EY and Deloitte.

The irony is KPMG’s website features an article titled “Essential Elements of Responsible AI: How Solid Guardrails Can Help You Scale AI Faster.” They’re selling AI expertise while failing at basic fact-checking.

David’s sarcastic take nailed it. “The only way I could think this could work is if the Big Four can go to the Fortune 500 and be like, ‘Look, we know all the mistakes that can be made. Now listen to us because we know what not to do.’”

How to Actually Use AI: A Real-World Success Story

While KPMG was publishing fantasy case studies, Blake was using AI to do real client work and showing what responsible AI use looks like.

He needed to complete two years of write-up work for a service business: 2,200 transactions across nine accounts, with source documents in a messy mix of PDFs and CSVs from different banks. In the old desktop days, this would have taken days of manual entry. Even with cloud accounting, it would take many hours of importing and coding.

Using Claude Cowork, he finished everything in about four hours, including gathering documents.

His approach was smart and deliberate. He pointed Claude at folders of bank statements and had it extract all transactions into Xero-compatible import files. It did OCR on PDFs, merged CSVs, and organized everything by account. Then he gave Claude the prior year’s general ledger and asked it to categorize transactions, but with a key addition: a confidence score for each categorization.

“I could open that up, sort by that score, and look at the transactions that are less than 90%,” Oliver explained. Instead of reviewing 2,200 items, he focused on exceptions.

The results were impressive. Claude missed just six transactions out of 2,200, and two of those were due to credit card statement date issues, not AI error. When a $5,000 clearing account discrepancy appeared, Claude opened Xero in a browser, analyzed the details, and identified the problems itself. One was a returned payroll miscoded to transfers. The other was more complex: undiscovered transfers to a business line of credit. Claude suggested this possibility, Oliver confirmed by pulling statements, and Claude then created the loan account, separated principal from interest, and fixed everything. That kind of discrepancy usually requires hours of investigation.

“I didn’t just say, ‘Here’s the GL detail, here are the transactions, go code them all and enter them into Xero,” Oliver emphasized. “I wanted to review it first, and I caught significant stuff.”

This capability is becoming more accessible. Microsoft’s Copilot Cowork is now available, with over half of Fortune 500 companies trying it during preview. Microsoft says it’s 30-40% cheaper per prompt than Claude, and since most accounting firms use Microsoft 365, it might already be on your computer.

Not everyone’s getting it right, though. David shared his frustration with QuickBooks AI. When he uploaded a PDF containing 12 monthly bills, QuickBooks mashed them into a single bill with line items from each invoice. No questions asked.

“It should say, ‘Hey, I noticed there are 25 bills in here. Do you want one bill or 25 separate bills?’ And I would just answer,” David said, comparing it to AI coding tools that ask before acting.

The Market Is Already Picking Winners and Losers

Meanwhile, CPA firms are seeing interesting pricing patterns. According to CPA Trendlines, overall pricing is up 4.2% year-over-year, reversing last year’s decline. But looking more closely at the breakdown, tax prep and planning jumped by nearly 8%. Advisory work rose over 6%. Audit only increased by 2.3%.

Clearly, clients will pay more for services requiring human expertise and judgment. Tax planning and advisory command the biggest premiums. More routine, standardized work, like audit, lags behind.

“Clients are willing to pay for tax planning advisory, for the human in the loop to make sure that the numbers are right,” Oliver said. “AI isn’t putting pressure on those fees at this point. And I don’t expect it to.”

The Real Divide: Verification vs. Vibes

The stories from this episode are different views of the same shift. KPMG publishes an AI report that’s 89% wrong while a solo practitioner uses AI to finish two years of work in an afternoon with near-perfect accuracy. CPA firms raise tax planning fees by 8% because clients value human judgment.

AI compresses the value of routine, unverified work while amplifying the premium on carefully applied expertise.

The divide in professional services is between those who verify and those who just publish. Between practitioners who build workflows with confidence scores and exception review, and firms that let AI-generated content sail through with fake citations. Between organizations that treat AI as a force multiplier for human expertise and those that use it to substitute for expertise they never had.

The practical takeaway is to learn the tools, whether that’s Claude Cowork, Copilot Cowork, or whatever comes next. But build human checkpoints into every workflow. Use confidence scoring. Review exceptions. Don’t set it and forget it. The productivity gains are potentially five to ten times traditional methods, but they disappear the moment you skip verification.

The market is already pricing this reality. Clients pay more for advisory, planning, and the assurance that a qualified human reviewed the work. Firms and practitioners who master this balance will command premiums. Those who don’t will find themselves on the wrong side of a restructuring that’s happening right now.

To hear Blake Oliver’s complete breakdown of his AI workflow, David’s full critique of accounting software AI, and more details on KPMG’s “vibe citation” disaster, listen to episode 493 of The Accounting Podcast.

A controller at a SaaS company that processes $50 million a month through its marketplace went on a two-week vacation. When he returned, his AI agents had already coded, categorized, approved, and synced 2,000 transactions. He reviewed just 67 (about 3%) by hand, and the entire cleanup took 30 minutes.

James Agius, Financial Controller at Skool, described his actual workflow on a recent episode of The Accounting Podcast. And it landed alongside benchmark data proving that, for the first time, off-the-shelf AI models from OpenAI, Anthropic, and Google are outperforming human accountants at basic bookkeeping tasks.

Hosts Blake Oliver and David Leary unpacked a series of developments that signal a genuine turning point for accounting. New studies from Digits and Ramp put hard numbers on AI’s bookkeeping abilities. A venture-backed startup led by a former PCAOB board member is building an AI-first audit firm. And KPMG’s entire US management committee flies to Silicon Valley every five to six weeks to meet with startups it views as potential threats.

But AI isn’t arriving to replace a surplus of accountants. It’s showing up amid a talent crisis that has more than tripled the number of unfilled accounting roles in a single year.

The Numbers Don’t Lie: AI Now Matches Human Bookkeepers

For years, the accounting profession has heard promises about AI. Now there’s data to back them up.

Digits just released the fourth version of its benchmark study, and CEO Jeff Seibert shared the results in an interview with David, which is featured on the episode. The test included categorizing over 2,000 transactions across multiple businesses into the correct chart of accounts. They tested all the major AI models (OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini) against outsourced human accountants.

“All of the major model providers have, for the first time, beaten real, outsourced human accountants at bookkeeping tasks,” Jeff told David. The humans scored about 79% accuracy. The AI models came in between 79.4% and 80.7%. The margin is small (about 1.6%), but the direction is clear.

Before anyone dismisses 79% as a low bar, Jeff offered important context. That’s actually typical for outsourced accountants who understand general accounting principles but don’t know the specific business. “They don’t know anything about that business or its industry, supply chain, geography, or customer base,” he explained. That missing context accounts for the 20% error rate.

What’s striking is how similar all the models performed. They’re all within three percentage points of each other. As David put it, basic transaction categorization “is kind of a commodity now.” It’s something everyone will essentially get for free from these models right out of the box.

But purpose-built systems go much further. Digits’ own AI, which learns from each business’s transaction history and can’t hallucinate by design, hits 97.8% accuracy. “Digits mimics the knowledge of a dedicated accountant who you’ve worked with for a number of years,” Jeff said.

The picture changes when you look at more complex work. Ramp tested its new Stack platform on 237 accounting tasks across eight synthetic businesses for categorization and financial close work. Its system scored 65.8%, beating the raw models but well short of perfect. This matches what most accountants experience. AI is great at pattern recognition but still struggles with judgment-heavy tasks.

AI still falls short in complex accruals, according to Jeff. Journal entries, fixed asset schedules, and prepaid expenses are the remaining frontier. Digits responded by launching automated accrual schedules where the AI identifies potential prepaids or fixed assets, drafts the schedule, and the accountant approves it.

Jeff drew an interesting parallel. At his tech company, engineers went from zero AI use to 100% in a single quarter. Jeff himself hasn’t written code since December, despite coding being his passion since age 12. “We have not fired our software engineers,” he said. “They are still critical, but the day to day has changed completely. Instead of them writing the code, they’re guiding the agents.”

One Controller, Zero Staff, $50 Million in Monthly Transactions

James Agius proves what these benchmarks mean in practice. He’s the financial controller at Skool, a SaaS company running online educational communities. The company handles over $5 million in monthly spend with nearly $50 million flowing through its marketplace each month.

James is also the company’s entire finance department. The company doesn’t have any staff accountants, AP clerks, or analysts. It’s just him and seven specialized AI agents, plus an eighth admin agent that checks the others’ work and enforces controls.

When Agius took two weeks off, those 2,000 transactions piled up. His automations handled almost everything, from coding, categorizing and approving to syncing to the ERP. When he returned, just 67 transactions needed human judgment. The cleanup took 30 minutes.

“His job changed from doing the work to reviewing the work,” Blake explained on the podcast. That shift freed Agius for forecasting, cash management, and strategy. It’s the work finance leaders always say they want to do but rarely have time for.

The timing couldn’t be more ironic. Just as AI enables one person to run an entire finance function, the profession can’t find enough people to fill open roles.

A Personiv study cited in Accounting Today found that the number of unfilled accounting and finance positions per company jumped from 5 to 17 in a single year, more than tripling. Eighty-four percent of finance and accounting leaders say there’s a talent shortage. The hardest role to fill is the senior accountant role, cited by 43% of respondents.

The drivers aren’t mysterious. The profession has talked for years about how 75% of CPAs were approaching retirement. “Well, now they’re doing it,” Blake said. And the pipeline is thin because staff accountants have been leaving after just a few years.

As David pointed out, senior accountants are exactly the people who would manage AI agents, so the talent shortage and the AI transition are colliding at the worst possible moment.

Firms are responding by racing to adopt AI. Sixty-three percent of leaders use AI to ease hiring pressure, up from 23% last year. For example, Bennett Thrasher moved talent acquisition from HR to the growth function, treating recruiting as strategically as business development. “The human labor becomes more valuable because it’s augmented,” Blake noted.

The Race to Reinvent

The competitive landscape is shifting as fast as technology. New entrants and incumbents alike are making moves that suggest they see this transformation as irreversible.

Christina Ho, former PCAOB board member and past podcast guest, joined Oath, a venture-backed firm building an AI-native audit practice from scratch. No legacy systems or technical debt. It’s AI-first from day one. They raised $6.6 million in seed funding and aim to automate 80% of audit work by 2030.

Oath plans to connect directly to clients’ accounting systems for continuous verification rather than year-end evidence gathering. CEO Lucas Ward emphasized audit remains “a human accountability function” even as machines handle verification. They’re recruiting “accounting engineers,” hybrid roles combining accounting expertise with computer science skills.

The Big Four are taking notice. KPMG’s US CEO now takes the entire management committee to Silicon Valley every five to six weeks, meeting with venture firms like Andreessen Horowitz and Bessemer to identify potential disruptors. They’re open to partnerships or investments, anything to avoid being blindsided.

On the platform side, Ramp’s new Stack product shows where AI agents might actually live in the workflow. Stack connects to existing tools like QuickBooks and accepts plain-language instructions, like “This client allocates revenue by location, not department. Split it across six cost centers.”

As Blake observed, “The GL is not the best place for agents to live. You want the agents at the point of the transaction.” Ramp already sits at the point of spend, giving its agents rich context about each business. The market agrees. Ramp just raised $750 million at a $44 billion valuation.

Not every AI adoption strategy works, though. KPMG rolled out a dashboard requiring employees to use AI for roughly 75% of their working time. Predictably, employees immediately gamed it. They had AI summarize emails they’d already read or generate random drawings — anything to hit targets. Blake called it “token maxxing,” comparing it to padding billable hours. Amazon shut down a similar program after seeing the same behavior.

What Humans Still Own

Where does human value go when AI handles the routine work? Jeff identified three things AI can’t replace.

1. Judgment. “AI goes off in weird directions,” he said. Experienced professionals must guide it through ambiguous calls.
2. Trust. “The AI will tell you anything you want. You can never trust AI.”
3. Accountability. “It’s never going to be liable for the numbers it gives you. What are you going to do, sue your AI?”

These are the differentiators for accountants who want to stay relevant as machines take over the rest.

All of the evidence from this episode points to AI crossing the competence threshold for basic bookkeeping and advancing toward complex tasks. One controller already runs a $50 million operation solo. Yet unfilled roles have tripled. Senior accountants are impossible to find. The retirement wave is here, and the pipeline is thin.

To thrive, you need to bring what AI can’t: judgment, trust, and accountability. The transition is here.

Listen to the full episode for the rest of Jeff’s interview, details on KPMG Australia’s whistleblower scandal fallout, and a discussion of the IRS leadership vacuum.