One exists that’s worth the time. Stanford’s Jung Ho Choi and MIT Sloan’s Chloe Xie spent roughly a year studying real AI adoption in accounting — not surveys about intentions, but actual behavior. They surveyed 277 practicing accountants, then partnered with an AI accounting software company to analyze transaction-level data from 79 small and mid-sized firms, covering hundreds of thousands of real transactions. This is closer to a randomized field study than anything else currently published on the topic. Here’s what it actually found, and what it means if you run a small firm.

Finding 1: AI didn’t replace the work, it moved it

The most important number in the paper isn’t a headline efficiency stat. It’s this: accountants using AI shifted roughly 8.5% of their time away from routine data entry and toward higher-value work — business communication and quality assurance, specifically.

That’s not “AI does the work now.” It’s “the work changed shape.” Nobody in this study got to do less work. They got to do different work — less typing, more judgment. If you’re evaluating a tool by asking “will this let me cut hours,” you’re asking the wrong question based on what actually happened in these firms. The better question is: “will this let my team spend the hours they already have on something that pays more per hour?”

Finding 2: adopters serve meaningfully more clients

Accountants using AI supported roughly 55% more clients per week than non-users in the same firms. That’s a real capacity number, not a vendor promise — and it lines up with the time-reallocation finding above. Fewer hours on data entry per client means more capacity for additional clients, without proportionally more headcount.

For a 5-person firm, this is the actual business case, more than “efficiency” as a vague concept: it’s a capacity lever. If you’re trying to grow revenue without hiring, this is the mechanism, and it’s backed by transaction-level data rather than a case study a vendor cherry-picked.

Finding 3: senior and junior staff use AI completely differently — and this is the part firm owners are missing

This is the finding that should actually change how you roll out any AI tool internally. Senior accountants treat AI outputs the way they’d treat a junior staffer’s first draft: as something to check. When the tool flags low confidence in its own output, senior staff push back, dig in, and apply judgment.

Junior accountants do the opposite. They’re more likely to accept AI output at face value — even when the tool itself is flagging uncertainty. This isn’t a generational attitude problem. It’s a repetition problem: senior staff know what a correct answer looks like because they’ve produced thousands of them by hand. Junior staff often don’t have that pattern-matching yet, so they have less basis to notice when something’s off.

Practical implication: if you’re rolling out an AI tool at your firm, the review process can’t be “the same tool, less senior oversight.” The people who need the most oversight built into their workflow are exactly the people you’re most likely to hand the tool to first, because their time is cheaper. That’s backwards from what this research suggests is safe.

What this doesn’t tell you

The study is honest about its limits, and so should you be: it’s early evidence from one partner firm’s software and one snapshot in time, not a universal law of AI in accounting. It also doesn’t address the judgment-heavy advisory work that’s driving most of the current hype — the data here is about transaction-level, task-based work, which is exactly the category where the previous issue’s framework said the real payoff tends to live anyway.

The one-sentence takeaway

If you’re rolling out AI at your firm: expect it to change what your team does, not just how fast they do it, expect the real ROI to show up as capacity rather than saved hours, and build extra review into junior staff workflows specifically — not less.

Here’s what’s actually in the market, sorted by what they are rather than how they market themselves, plus the one honest signal I could find underneath the noise.

Two categories, not one

Almost everything marketed as “AI bookkeeping” falls into one of two very different buckets, and firm owners frequently evaluate them against each other by mistake.

Software you or your team operate. Tools like CodeIQ, Booke AI, and Xero JAX sit on top of QuickBooks Online or Xero and handle transaction categorization and reconciliation. You’re still the one running the process — the AI just does the first pass. Pricing here runs roughly $18–$100/month per company, and these are built specifically for firms managing multiple clients.

Managed services with AI underneath. Bench, Pilot, Zeni, and Botkeeper’s enterprise tier are a different product entirely — a human bookkeeping team does the work, with AI handling the repetitive parts behind the scenes. You’re buying outsourced bookkeeping, not a tool. Pricing reflects that: $189–$549/month and up, sometimes tied to your transaction volume or revenue.

If you’re comparing a $20/month categorization tool against a $300/month managed service and wondering why one seems so much more expensive, this is why — they’re not competitors, they’re different purchases.

The one honest signal in the noise

Vendor blogs all claim accuracy numbers in the high 90s and dramatic time savings — one company’s own press claimed 99.97% accuracy, which is the kind of precise-sounding figure that should make you more skeptical, not less, since real-world bookkeeping accuracy is genuinely hard to measure that precisely.

Independent review platforms tell a messier, more useful story. On Trustpilot, Bench’s reviews split sharply: several clients rate it highly for reducing their workload, while others report significant frustration — one recurring complaint across multiple reviewers is bookkeeper turnover, with one client reporting they’d been assigned somewhere around 10–12 different bookkeepers over four years. That’s a real operational detail no vendor blog will mention, and it matters more than any accuracy percentage if continuity with your client relationships is the thing you’re actually protecting.

The takeaway isn’t “avoid Bench” or “avoid managed services.” It’s that the actual failure modes of these tools show up in independent reviews, not comparison blogs — and if you’re recommending a tool to clients or adopting one for your own firm, that’s where to look before you look at the pricing page.

How to actually evaluate one of these

Going back to the breakeven framework from last issue: the real question for any tool in either category isn’t the accuracy claim, it’s two things —

1. What’s the actual review burden once it’s running — not week one, but month three, once the tool has learned your chart of accounts or your client’s transaction patterns.

2. What’s the failure mode when it’s wrong — a mis-categorized transaction your team catches in review is a minor cost; a mis-categorized transaction that reaches a client’s financials because review got lax is a real one.

Software-you-operate tools tend to fail in the first way (annoying, catchable). Managed services tend to fail in the second way when they do (continuity and oversight gaps, like the Bench pattern above) because you have less visibility into the day-to-day process.

What’s next

I’m going to pick one tool from the software category — likely Booke AI or CodeIQ, given the lower cost of entry — and actually run it against a real (anonymized) dataset over the next few weeks. Next issue will have real numbers, not vendor numbers.

That’s the only question worth asking. Not “is this tool impressive,” but “does the math work for a firm that bills in six-minute increments.”

The real cost of adoption

Every AI tool has two costs, and vendors only ever talk about the first one.

The subscription cost is the easy number — usually somewhere between $20 and $300 a month depending on what you’re buying. Firm owners are good at evaluating this one. It’s the second cost that gets missed.

The adoption cost is the hours spent learning the tool, correcting its mistakes, building the review process around it, and — critically — the hours a partner or senior staffer spends checking its work until trust is earned. For most tools, this is 10 to 30 hours in the first month. At a partner’s billable rate, that’s not a rounding error. That’s real money spent before you’ve saved a single minute.

This is why so many firms try a tool, get frustrated in week two, and quietly stop using it. The math never had a chance to work, because they measured the tool against day one instead of month three.

Where the payoff curve actually bends

Time savings from AI tools in accounting work tend to follow a pattern: flat or negative for the first few weeks, then a bend upward as the review burden drops. The tools that pay off are the ones where that bend happens fast — because the task is narrow, repetitive, and has a clear right answer.

That points to a simple filter for any tool you’re evaluating:

• Narrow task, clear right answer (bank statement categorization, document data extraction, first-pass reconciliation flags) → adoption cost is low, payoff curve bends fast.

• Broad task, judgment-heavy (client communication, advisory recommendations, anything requiring context about a specific client relationship) → adoption cost is high, and the review burden may never drop enough to justify it.

Most of the AI-in-accounting hype is aimed at the second category, because it sounds more impressive in a demo. Most of the actual, measurable time savings live in the first.

A rough breakeven formula

You don’t need a data science team to run this. For any tool you’re considering:

Breakeven weeks ≈ (Monthly subscription cost + estimated adoption hours × your blended rate) ÷ (Weekly hours saved × your blended rate)

If a tool costs $100/month, takes 15 hours to properly adopt at a $150 blended rate ($2,250), and saves your team 3 hours a week once it’s running, at $150/hour that’s $450/week in saved time. You’re looking at roughly 5-6 weeks to breakeven — reasonable. If that same tool only saves 30 minutes a week, you’re looking at 30+ weeks, which is a much harder sell to a partner group watching the clock.

The formula is rough on purpose. The point isn’t precision — it’s forcing the adoption cost into the conversation before you commit, instead of discovering it in month two.

The uncomfortable part

Most firms currently evaluating AI tools are comparing subscription price against a vague sense of “efficiency,” not against this kind of breakeven math. That’s not a criticism — it’s just where the profession is right now, because most of the marketing around these tools is designed to keep the adoption cost invisible.

That’s really what this newsletter is for. Not more tool hype. A running, honest account — task by task, tool by tool — of where that breakeven line actually falls, so you’re not the one paying to find out.