“Anybody who says they’ve built autonomous AI in accounting or finance is full of crap. Nobody has figured out how to do that yet.”
That’s how Blake Oliver opens Episode 497 of The Accounting Podcast, and it sets the tone for the whole show. Blake and co-host David Leary spend the first half on tax policy and platform drama, then spend the second half talking to Sam Leon, founder of the Millennial CPA, a one-person, tech-enabled tax practice in Richmond, Virginia. Accounting Today named the firm to its 2026 Best Firms for Technology list.
Put the two halves together, and you get a clear argument that AI isn’t taking accounting work away. It’s moving it, and where it lands is professional judgment. Automating the prep stage pushes the bottleneck onto the scarcest people in any firm: the managers, controllers, and partners who have to review everything the machine produces. Three threads carry that case:
- The “verification tax” and what the jobs data really shows
- Why older platforms and workflows are much harder to replace than the market believes
- Sam’s practice, which proves the payoff comes from automating everything around judgment rather than the judgment itself
The “Verification Tax”: Why the Bottleneck Just Moves
The first thread starts with a simple problem. Someone still has to check the work. Blake points to reporting by Accounting Today technology editor Chris Gaetano, who found that AI’s promised productivity gains in finance are getting eaten up by the time it takes to check, explain, and govern AI outputs. A recent Sage survey found that nearly half of finance professionals spend more than 15 hours a week on verification, and 19% spend more than 30 hours. Sage calls this the “verification tax.” Only 9% of respondents plan to give AI broad control over transactional finance. The profession isn’t letting these tools run on their own.
“We’re just shoving the bottleneck to a different spot,” David points out. “The bottleneck in theory was the data entry. Now we’ve moved it to the review of the data.” Blake takes it one step further. Speed up the prep work, and you simply pile more onto the reviewers. And there aren’t enough qualified reviewers to handle it. As he says, “We don’t have enough managers and directors and partners. We don’t have enough controllers and CFOs.”
Blake is careful not to dismiss the tools because he feels AI sharpens his own judgment. “It allows me to make decisions faster, to figure things out quicker. But I still have to think a lot.” That thinking takes skill and experience, which is exactly why you can’t automate the review layer away.
The Jobs Data Contradicts the “AI Replaces Accountants” Story
If review is the real constraint, then firms should need more skilled people, not fewer. The data the hosts cite says they do. Research from Ramp and Revelio Labs tracked AI spending and workforce records at nearly 22,000 U.S. companies from 2021 to 2026. Firms that spent more on AI grew total headcount by an average of 10% in the two years after rollout. The heaviest investors expanded entry-level hiring by 12%.
David adds a report from Indeed’s Hiring Lab showing that mentions of AI in job titles and descriptions have more than tripled since 2022. On top of that, 63% of AI-titled roles now sit outside tech companies. AI is becoming a required skill in ordinary, nontechnical jobs. The hosts call this shift up-leveling. As Blake puts it, “The workers we need are higher level.”
Sticky Systems and Technical Debt: Why Xero and QuickBooks Aren’t “Toast”
The same durability argument applies to the software underneath the work. David walks through the drama at Xero. Investors in New Zealand and Australia are uncomfortable with the large pay package for its U.S. CEO. This week, she sold all her remaining shares for $2.2 million to cover a tax bill. The stock is down roughly 58% over the past year. All of it feeds a story that AI-native startups will bury the incumbents.
Blake thinks the market has it wrong. To believe that story, you have to believe small businesses will start coding their own accounting software. He tried it himself. “Yes, it’s doable, but the problem is then you have to review so much, and everything looks so good that it’s hard to know if it’s right.” You need rails, or a general ledger like QuickBooks or Xero, so that whoever handles the tax work can trust the numbers. Both hosts argue the incumbents could actually win, because AI removes the hardest part of building software: the user interface. Expose the ledger through MCP connections, and users can work through simple conversation while the trusted structure stays in place.
Then David shares a story that illustrates how “sticky” legacy technology can be. A Texas filtration company, Sparkler Filters, ran IBM’s 402 accounting machine (a punch-card system introduced in July 1948) all the way until 2020 because replacing it meant retraining staff, disrupting decades of process, and risking errors.
Blake turns that warning on AI itself. Workflows he built two years ago started breaking as models were retired and integrations changed, and he’s the only person at his company who can fix them. “You’re going to end up spending on a team that can maintain those tools. You are now a developer or an engineer.” Call it technical debt. It’s a cost almost nobody budgets for.
Sam Leon’s Firm Automates Everything Before Judgment
That brings us to someone who has built a whole practice around this idea. Sam Leon spent 13 years in tax before going solo. He left because being truly tech-enabled isn’t something you can get signed off on inside a 20-person tax department.
His first idea, a year ago, was to have AI agent A and AI agent B play different firm roles. He dropped it. The technology “was not quite there,” and it still meant a lot of copying and pasting. So he flipped the problem around. Before anyone enters a single number into a return, three to seven hours of work has already happened. Automate that.
It starts with a custom smart intake form designed to scope engagements accurately and avoid the chronic over- and under-scoping he watched at earlier firms. A good call leads to a templated engagement letter in Ignition, which kicks off automated billing. A Slack bot he built populates his CRM and opens a client profile in TaxDome. Claude generates the list of expected documents, the client portal opens, and documents flow in.
Next comes the AI preparer. Sam keeps a Claude project folder for each client, holding redacted documents, and runs a conversational, deliberately custom process that produces an Excel workpaper organized by schedule. Claude understands that there are hundreds of possible schedules. What it doesn’t understand is why a given number belongs in a certain place. So Sam gives it guardrails, or general guidelines for where investment income or a home sale should land.
The AI reviewer is the reverse, and he set it up as its own standardized project because a review runs the same way every time. It performs a three-way match: the current return, the prior-year return, and the pile of source documents. The documents answer “are the numbers right?” The prior year answers “did we miss something?”
The results are concrete. A C corp workpaper that used to take six or seven hours now takes about an hour of back-and-forth. A two-hour individual workpaper takes roughly 15 minutes. That’s a 5x-plus boost on the exact work that used to eat up tax season.
Sam is firm about the caveat, “Don’t try this at home unless you’ve been a tax preparer for a while.” He still keys the numbers into the return by hand, which doubles as a check. AI is a preparer and reviewer assistant, not a replacement.
The Economics: Price for Expertise, Not Hours Saved
If AI cuts the work in half, why not cut the fees too? Sam doesn’t. His $200-a-month Claude Max subscription is, as Blake puts it, a no-brainer against the hours it saves per return. And it isn’t even his biggest line item. Practice management, intake software, and the tools he tries out and cancels account for more of the roughly 70% of his budget that goes to technology.
He has never billed by the hour. He prices fixed packages by scope, complexity, and judgment, and he recently raised his minimum to $1,500 from about $1,250. A corporation with an international subsidiary and three international partners may take less time to key in, but “there’s still a lot of judgment, and there’s still a lot of professional experience going into it.” No client has asked for an AI discount.
His software, TaxWeave, follows the same logic. It consolidates client information from email, document portals, cloud storage, and team notes into a single view of where each client stands. It solves the “master Google spreadsheet” problem his old firms could never quit, even after buying practice management software, and layers agent actions on top. It attacks the chaos around the work, not the judgment inside it. Having reached the limits of what he can build as a self-described novice coder, he’s brought on a software engineer.
The Unglamorous Playbook
From the verification tax to punch-card machines to Sam Leon’s workpapers and price list, episode 497 keeps making the same point from different angles. AI relocates accounting work onto professional judgment rather than removing it.
The playbook is durable and distinctly unglamorous. Aggressively automate intake, organization, and comparison. Budget for growing review and maintenance burdens. And price for expertise rather than for the hours you just saved. Your value is in the trust and judgment layered on top of the data, not in the data handling itself.
Hear Sam’s full end-to-end walkthrough, plus Blake and David’s complete analysis, in Episode 497 of The Accounting Podcast.
