“One day, the board is going to ask the CEO, ‘I see you spent all that on tokens. What was the result?’ And they’re not gonna be able to answer it.” That’s David Leary on Episode 496 of The Accounting Podcast, and in one line he captures the tension running under nearly every story he and co-host Blake Oliver covered this week.
Here’s the big idea that ties all of this episode’s stories together is that, as AI lowers the barrier to building software, automating audits, and streamlining everything from month-end close to CPE reporting, the accounting profession’s real value shifts away from doing the work and toward overseeing it. That’s because every advance comes bundled with a hidden cost or risk. The real professionals know exactly where the practical controls, real costs, and actual risks live.
Xero Arms the “Builder Class,” but It’s Still Very Early Days
Xero turned 20 this year, and at Xerocon London the company used that milestone to lean into what CEO Sukhinder Cassidy calls a “builder class” movement. The builder class includes accountants, bookkeepers, and business users who can create automations and software-like tools without being traditional developers. As David noted, this connects directly to Xero’s recent developer-channel push, and its Is Everyone a Developer Now? YouTube channel that once seemed puzzling but now reads as strategy.
The centerpiece is XeroForce, an invite-only, alpha-stage, no-code AI agent builder. David compared it to Zapier because you connect your apps, describe a workflow in plain language (“every time an email like this comes in, pull the PDF attachment and post it as a bill”), and it builds the automation for you. It’s part of Xero’s broader play, alongside the new mid-market Xero Ultra product, to keep customers on its full stack as they grow rather than losing them to a Sage Intacct or Oracle NetSuite.
Since Xero started connecting Claude and other agents in January 2026, its API usage jumped 400%, and 2,000 customers connected Xero to Claude in the first 60 days after the announcement. That sounds impressive until you do the math. Against roughly 5 million Xero businesses, 2,000 is about 0.04%. “We’re so early still,” Blake said.
David’s advice was blunt: “Don’t get FOMO, because the number of people actually doing it is so, so teeny, teeny, teeny.” He also flagged the missing “database layer.” Accountants can vibe-code an app, but there’s often nowhere for it to live and no easy way to host and maintain it. That’s the practical control problem hiding behind the promise.
Vibe Coding: Real Six-Figure Savings, Real Key-Person Risk
If Xero arms accountants to build, some firms aren’t waiting for a vendor at all. As one listener put it, “Small firms can now develop their own software for less than the cost of buying software.” For example, at HoganTaylor, Randy Nail’s team needed a financial reporting tool for a new audit method. A vendor quote came in around $200,000 a year. Instead, they built it themselves in Excel with AI. Blake says that’s the smart way to do it: build in a familiar tool you already own, not a standalone app “living somewhere on a server” you don’t fully understand. Mike DeKock of MJD Advisors went further, replacing $300,000-a-year audit software with a build in Claude and Retool for under $30,000 annually. Even Starbucks is chasing the same impulse at enterprise scale, trying to trim its roughly $400 million software spend by building more in-house.
The catch in DeKock’s case is that he’s the only one who knows how it works. That’s key-person risk. Blake and David have lived it. Earmark’s AI course generator runs on about 75 Zapier steps Blake built years ago, and when it broke recently, only Blake could fix it. It’s the same problem as the notoriously complex financial model that only one person understands, where everyone else is afraid to click the wrong cell.
David pushed back, and fairly. AI coding tools document their own work well. It includes comments in the code and plain-English summaries, so it may be less of a problem than it first appears. And vendors aren’t a guaranteed safety net either. He recounted a Streamyard support headache: “If I have to use an AI bot to get support for your product, I might as well just chat with a different AI bot and have it build me a replacement.” Still, Blake summarizes, “You’re saving money now, but you’re creating risk potentially in the future.” Build where you’re comfortable, and keep a backup.
Checkbox Compliance vs. Real Protection
That same tradeoff between what looks safe on paper and what actually holds up runs straight through this episode’s audit and security stories. MindBridge submitted formal comments urging the PCAOB to clarify how auditors should document, assess, and defend AI-assisted work, especially now that software can test an entire population of transactions instead of a sample. The existing standards were built around sampling. They simply don’t address risk scores, investigation thresholds, or what counts as “sufficient evidence” in full-population testing. Firms run the new AI-assisted procedures alongside the old manual tests because, as David put it, they “need the check box” to pass inspection.
Meanwhile, the PCAOB voted unanimously to seek comment on easing parts of QC 1000, the 2024 quality-control standard. The changes could remove the external quality-control function for firms that audit more than 100 issuers, and relieve registered firms that don’t actually perform PCAOB engagements. The hosts recognize the need to modernize but question whether simply rolling back standards is the answer. Blake framed the core issue as an inputs-based approach to regulation, not an outcomes-based one. “Having a system doesn’t necessarily mean that your audit is going to be quality.”
David’s parallel nailed it. A cybersecurity audit of 275 Australian accounting firms found 76% had no protection against email spoofing, yet nearly all almost certainly have a required written information security plan (WISP). “You don’t have to be secure,” David said. “You just have to have a plan.” Or, more pointedly, “You must spend time building this document about your security plan instead of investing that time and resources into actually being secure.”
The Token Problem: Usage-Based Pricing and a New Kind of Cost Accounting
If compliance is about knowing where the real controls live, the next challenge is knowing where the real costs live. AI is rewriting software economics, from predictable, per-user subscriptions to variable, usage-based token spend. Tools like Claude Cowork can do far more now, but they cost more too. You ask an agent to do one thing, and it chugs away in the background, burning tokens and blowing past your allotment fast.
A KPMG survey of over 2,000 senior leaders across 20 countries put numbers to the pain. Only 29% feel they understand operating costs as they scale enterprise AI. Roughly a third cite limited understanding of AI economics as a barrier to deploying agents. And nearly half of organizations have re-phased AI deployments when costs exceeded expected value. Blake noted this could be “the next great area of cost accounting,” measuring where tokens get burned and proving where they deliver results.
Part of the answer is matching the model to the task. Lower-cost, high-fidelity models were the fastest-growing influence on AI strategy, up seven points from the prior quarter. Think Claude Opus versus Sonnet versus Haiku. Pick the right one not just per workflow, but per step within a workflow. Emerging “routers” now automatically direct each task to the most cost-effective model because, as David said, “you don’t need to blow $100 for a $0.02 answer.”
The Accountability Layer
Pull the threads together, and the pattern is unmistakable. Xero’s builder class and XeroForce, vibe-coded software, AI-assisted audits, token economics. Every advance in this episode came paired with a counterweight. Adoption is still a rounding error. Custom builds create key-person risk. Modernized standards risk hollowing out real oversight. And usage-based pricing makes costs genuinely hard to forecast.
Even the episode’s feel-good story carries a caution. One listener used Claude Cowork to handle Florida’s tedious CPE reporting, consolidating roughly 80 certificates into an Excel list, entering them in the state portal, and even catching and fixing its own duplicate entries and combining PDFs into a single upload. It was a “relatively low-risk” win, done with his own data while he caught up on Severance. But David’s cautioned listeners to check the portal’s terms of service, since older, pre-AI site terms may prohibit bots.
The need for judgment is the through-line. As AI collapses the barrier to building and automating, doing the work stops being where accountants create their edge. The durable value moves to oversight. You need to know where the practical controls, real costs, and actual risks live. The tension is always between innovation and accountability, and accountants are uniquely positioned as the accountability layer. So start experimenting, but build where you’re comfortable (Excel is fine), keep a backup for anyone building custom tools, and start measuring your AI spend now.
There’s plenty more in episode 496, including Lionel Messi’s roughly $28 million potential U.S. tax bill and FIFA’s tax-exempt status, the activist-investor fight over CBIZ’s acquisition strategy, and the full World Cup betting-tax breakdown. Listen to the whole back-and-forth on The Accounting Podcast, and don’t forget you can earn free NASBA CPE for the episode through Earmark.