If you strip out the money pouring into artificial intelligence, the US economy “may have entered recession territory.” That’s the conclusion of the forecasting firm Oxford Economics, and it’s how Blake Oliver opened Episode 510 of The Accounting Podcast with co-host David Leary. Accountants should pay attention, because it means firms aren’t yet capturing the AI boom’s productivity gains that are propping up the economy. AI supercharges individual accountants. But Blake and David argue that firm-level payoff depends on standardizing work, controlling token spend and treating agents like permissioned employees.
The Economy Is Betting on AI ROI
Blake spotted the story on CFO.com. Michael Pearce, chief U.S. economist at Oxford Economics, says, “All the growth in the past couple of years has been in AI. Everything else has been falling.” By Oxford’s math, the economy grew 2% to 2.5% in the first half of the year and should grow north of 3% in the third quarter. But take AI out, and growth is roughly flat, possibly negative.
The irony is hard to miss. Pearce made these remarks at the CFO Leadership Council’s fall conference, where vendors were selling AI-enabled products.
“Have we actually started to make more money for our firms as a result of AI yet?” Blake asked, turning the question on the profession. At Earmark, his AI sales agent helped bring in podcast sponsorships. Overall, though, it added maybe “a few percent” that “could have happened anyway.” His verdict is “We have to see actual productivity gains to justify these billions or trillions of dollars of investment.”
Where AI Already Delivers, One Person at a Time
Some gains are happening. Blake pointed to a Zapier blog profile of Damian Guzman, a solo lawyer who caps himself at about 30 billable hours a week. After rebuilding his practice around Claude, Guzman estimates he gets roughly five times the output. Every hour, Claude scans his Gmail for client requests, creates tasks in Clio, his practice management software, and starts drafting. By the time Guzman opens a request, “roughly half the work is already complete.”
There’s a catch for pricing. Guzman switched to flat rates because, as Blake noted, if he kept billing hourly, “he’d have an 80% revenue decline.” Firms that offer client accounting services (CAS) faced the same challenge years ago.
Blake runs his own scheduled “email sweep.” It reads his inbox, drafts replies in his voice and checks his Google Calendar to suggest meeting times. These tasks work because they don’t need much background. As David put it, you can often reply “based on the context clues in that email.”
The Token Bill Comes Due
The more context an agent needs, the more it costs. Blake trained his system with growing files of operating rules, and every scheduled task reloads them. He’s on Claude’s $100-a-month plan and regularly runs out of tokens halfway through the week. “I’m at $2,500 for October, so I might spend like $3,000 in tokens.”
“We’re burning cash to reply to other AI agents,” David joked.
Blake weighs that against hiring. Is around $3,000 a month in tokens worth it to avoid a $70,000 to $80,000-a-year hire you’d also have to manage? That’s the math he’s running. But costs keep climbing for a structural reason. “It’s like your memory is erased and you have to go and read everything about that client,” Blake explained. Humans, he noted, consolidate what they learn while they sleep. AI stores context but never truly remembers it, so it rereads everything, every time.
Why Individual AI Doesn’t Scale to the Firm
This is the heart of the problem. Blake says AI “makes individuals with experience so much more productive. But it’s not eliminating any jobs.” Agents still need an experienced human to guide and review the work.
David compared today’s agents to a firm where employees never share knowledge. Each one works heads down, and when a client moves to someone else, it’s like starting over. “The firm-wide AI and the individual AI are completely different,” Blake said, drawing a clear line. That’s why investors get wowed. “The demos are so impressive because they’re not running at scale.”
The fix is old wisdom. When every client runs on a different system and process, onboarding a human takes forever, and the same is true for agents. “The tech isn’t magical in that respect. Not yet anyway.”
Orchestration layers that manage spend and shared knowledge are coming, including Ramp’s Glass and Pilot’s Meridian Starter. David thinks Intuit and Xero probably missed the opportunity while building collection-letter agents: “I need a protective layer before data goes into my GL.”
Private equity may be the proving ground. Platform Accounting Group just raised $355 million. Labor costs often make up 40% to 60% of a firm’s expenses, and Blake argues cutting those costs by even 10% to 20% would lift margins. And as David noted, standardizing small firms on cloud tools could get them partway there with no AI at all.
Treat Agents Like New Hires, Especially Near the Money
Security is the other precondition. Big banks, including JPMorgan Chase, Bank of America and Wells Fargo, restrict third-party AI agents because, David said, they “can’t tell the difference” between authorized agents and rogue ones.
Blake traced the problem to many banks never building multi-user permissions. Business owners already shared a single login with their accountants. Now they share it with agents.
His fix is to give Claude its own limited login, as he did in Xero, with the same access a bookkeeper would get. “We have to think about the AI as just being like another human team member.” Otherwise, David warned, “If you let AI use your login to our systems, now AI has God-mode for all our systems.”
The Payoff Is in the Boring Work
AI’s payoff for firms (and maybe for the economy) depends on unglamorous work like standardizing processes, managing token spend and setting permissions. Firm leaders need ROI, not impressive demos. Listen to Episode 510 for Blake and David’s full discussion, including Stripe’s push toward becoming an accounting operating system.