An off-the-shelf AI agent opened two general ledgers, compared the balances with prior-year returns, built book-to-tax workpapers in Excel, and entered two business returns into TaxAct. It even found and corrected errors in the tax software.
The process took a few hours. Yet the most important step was still human. Blake Oliver downloaded the final returns, checked them against the general ledger, and transmitted them himself.
The tension between machines doing the work and people remaining accountable runs through Episode 504 of The Accounting Podcast. Blake and co-host David Leary discuss agentic tax preparation, AI-native ledgers, flawed do-it-yourself cost segregation studies, accounting salaries, staffing, and failures of judgment at the Big Four.
AI Is Moving From Research Assistant to Tax Preparer
Last year, Blake mainly used AI as a research partner. This year, he gave Claude Cowork a much larger role in preparing two returns: his S corporation, which uses Xero, and Earmark Media’s partnership, which uses QuickBooks.
Claude pulled the profit and loss statement, balance sheet, and trial balance for each entity. It then compared those reports with the prior-year returns. The S corporation tied out, but the QuickBooks file didn’t. Claude traced the problem to a dropdown issue that caused reports to run on the accrual basis even though they appeared to be set to cash basis.
After resolving the mismatch, Claude asked Blake about book-to-tax adjustments and built Excel workpapers. It then opened TaxAct Business, completed the interview, entered the data, and worked through the software’s error checks.
The agent also found several problems. TaxAct omitted fully nondeductible entertainment expenses from its calculation, so Claude created a custom Schedule M-1 add-back. It corrected distributions that had defaulted to zero, fixed ending retained earnings so Schedule L balanced, removed an incorrect name-change selection, and addressed an Arizona filing checkbox that kept resetting.
David described the workflow as using AI to “bridge the gap between the GL and the tax return.” Blake estimated active agent time at about an hour per return.
That experience raises the question: if AI can connect the books and tax return, will those systems eventually become one platform?
AI-Native Ledgers Could Change the Technology Stack
David pointed to Accrual, an AI tax platform, agreeing to acquire Puzzle’s accounting-firm business and technology. General Catalyst backed both companies, so David viewed the deal as a possible combination of related investments.
Blake saw practical value in the pairing. Puzzle’s AI works within Puzzle, but accounting firms can’t move every client away from QuickBooks, Xero, and other ledgers. Accrual can let agents work across existing systems while providing a tighter experience with its own ledger.
A livestream discussion also raised the possibility of pairing AI with an open-source ledger. For basic write-up work, an accountant might give an AI tool a year of bank statements, create an import file, and avoid paying for features the client doesn’t need. Blake and David treated this as an idea worth testing rather than a proven replacement for established systems.
But the limits of general-purpose AI are clear when the work demands specialized evidence.
Polished Output Isn’t Defensible Work
Blake cited an Accounting Today article by Heidi Henderson of Engineered Tax Services. Over six weeks, three prospective clients brought her firm cost segregation studies created with ChatGPT or Google Gemini and asked the licensed engineering firm to validate them.
One Gemini spreadsheet contained six rows and assigned $100,000 of bonus-eligible basis to a $400,000 property. A ChatGPT workbook reclassified $1.24 million of a $4.7 million athletic facility to shorter-life assets.
Measured against 13 principal elements in the IRS audit technique guidelines, the ChatGPT study satisfied one, and the Gemini study satisfied none. They lacked items like a stated methodology, an engineer of record, engineering takeoffs, reconciliations, and required statutory analysis.
Blake argued that a specialized model trained on a firm’s methods and IRS guidance might prepare parts of a report. It still couldn’t perform the site visit, take photographs, or provide an expert’s sign-off. General-purpose tools, he warned, can “BS their way through it” and produce weak work that looks convincing.
That makes experienced review more important, but the profession may not be investing enough in the people who provide it.
Firms Need People Who Can Challenge the Machine
EY plans to distribute $100 million in bonuses tied to “human skills,” including adaptability, innovation, and judgment. Awards can reach $25,000 for material contributions, including team awards.
David viewed the program as paying employees to remain the human in the loop. Blake called it a smart way to align staff incentives with firm risk. Both hosts noted that all four Big Four firms have faced problems involving fabricated AI citations in published reports.
Yet median entry-level accounting pay fell from $75,000 to $73,000 even as compensation rose overall. David warned that firms especially need seniors and managers. These are the people who “watch the agents.” Lower starting salaries could weaken the future supply of those experienced reviewers.
Accountability Still Reaches the Top
KPMG Australia cut 360 employees and 27 partners after consulting revenue fell nearly 17%. The cuts follow allegations involving client data leaks, confidential client information used to win new work, and poor whistleblower investigations.
“Accountants will lose their jobs if people at the top-end management level are not doing ethical things, or if the firm’s culture is not ethical,” David said, pointing out the consequences.
Deloitte, meanwhile, agreed to pay $21.5 million to settle Justice Department allegations involving race- and sex-based employment practices on federal contracts. Deloitte denied the allegations and admitted no liability. A Hong Kong court also refused to dismiss PwC International from the Evergrande liquidators’ multibillion-dollar lawsuit, although the ruling didn’t decide the cases’ merits.
Technology changes how quickly work gets done, but it doesn’t change who must answer for the result. To benefit from AI, firms need to know how to direct it, test it, and take responsibility for its output.
Listen to the full discussion in Episode 504 of The Accounting Podcast.