Is AI ROI Real for CPA Firms? The Math, Not the Pitch

You've sat through the pitch. A vendor puts up a slide with a big ROI percentage from somebody else's firm, a payback timeline measured in months, and a case study with a name you don't recognize. The question you actually have is simpler and more honest than anything on that slide: will this pay back for a firm my size, or am I about to spend money I don't have on a system that's really just a bet on someone else's sales pipeline?

AI ROI for CPA firms is a real question with a real answer. It's just not the answer that anyone selling you a tool has an incentive to give you straight. Every number you've seen so far came from a vendor, an implementation shop, or a "free AI readiness assessment" that ends in a proposal, and none of them are neutral on the outcome. This piece is written by someone who isn't selling a CPA-specific software seat, so there's no reason to round the math in either direction. That's what follows below: the actual formula, a full worked example with honest numbers, and the return volume below which the math genuinely doesn't pay back for a small firm.

Here's what matters most, before the deep dive:

• The reusable ROI formula (value created minus total cost of the automation, divided by total cost) works for any tool, any vendor, any firm size.
• Licensing-only cost tracking, just the subscription line, commonly understates real automation cost by 40-60% once implementation, maintenance, and staff review time are counted.
• MIT research found roughly 95% of generative AI pilot projects fail to show measurable ROI. IBM found only about 25% of AI initiatives deliver the ROI they were expected to. Both are general enterprise findings, not accounting-specific.
• Below a certain return volume, custom automation for a small firm genuinely does not pay back in year one, and knowing that number before you sign is the reason to do the math yourself.
• None of this is about running with fewer people. It's about knowing what the freed-up hours are worth before you decide what to do with them.

What the "AI ROI for CPA firms" conversation gets backwards

Search this exact phrase and every result on page one is either a software vendor with a CPA-specific SaaS seat to sell or an implementation agency with a paid assessment to sell. That's not an accident. It's how the category is built: a vendor's case study is drawn from the best of their install base, not the median customer, and a payback timeline in a sales deck assumes the tool gets adopted correctly on day one by staff who need zero ramp-up. Almost none of them count the cost of the hours your own team spends learning the tool, reviewing what it produces, and fixing what it gets wrong.

This isn't a hunch. It shows up in research nobody was trying to sell you anything with. MIT's widely cited "State of AI in Business" work, the study people now call the GenAI Divide research, found that roughly 95% of generative AI pilot projects at companies fail to show measurable ROI. IBM's own research on enterprise AI initiatives found that only around 25% of them deliver the ROI they were expected to. Neither study is about accounting firms specifically. Both are general enterprise findings, from companies with far bigger budgets and dedicated IT staff than a four-partner CPA practice. If ROI is this shaky industry-wide, an accounting-specific ROI number in a vendor's deck is the exception being sold to you as the rule.

There's a reason a Forbes Councils piece on small accounting firms and AI specifically flagged that smaller practices carry more AI risk than large ones, not less. A large firm can write off a failed pilot as a rounding error and try again next quarter. A four-partner firm usually can't afford that mistake twice. If you're the one signing the check, that asymmetry is exactly why you should want the real math instead of another testimonial. Most AI tools for CPA firms are priced and marketed to keep that conversation short, because a long one invites the buyer to run the math this article is about to walk through.

How to measure AI ROI in accounting: the actual formula

Strip away the vendor language and the formula is not complicated:

ROI = (value created − total cost of the automation) ÷ total cost

Value created is the honest dollar value of the staff hours the automation frees up, priced at what that time is actually worth to your firm: a billable rate, room for more return volume without adding headcount, or time redirected to advisory work. Not a vague "efficiency gain." A number you'd defend in front of a partner.

Total cost is where almost every vendor's math quietly breaks. It has to include everything the automation actually costs, not just the invoice: the build or implementation, ongoing maintenance, and the piece that never makes it onto a sales slide, which is the staff hours spent reviewing what the system flags and correcting what it misses. Add up subscription line items alone and you'll land on a number that flatters the vendor. Licensing-only cost tracking like that commonly understates the real total cost of an AI or automation project by 40-60%, because implementation, maintenance, and exception-review time never get counted. That gap is exactly why a properly scoped CPA tax document automation engagement prices in maintenance and review time from the start, instead of leaving a firm to discover those costs on its own in the middle of a busy season.

Once you have both numbers, value created and total cost, counted honestly, the formula does the rest. No case study required.

The worked example: a 4-partner firm, 800 returns a season

Here's the math, run through with round, hypothetical numbers. This is an illustrative scenario, not a real firm. Plug in your own numbers and the formula holds exactly the same way.

Say a 4-partner firm processing 800 returns a season automates document collection: chasing missing paperwork, following up, and organizing what comes back, the work that eats staff hours from January through April. The total cost of that automation for the season, including build, maintenance, and the senior-staff time spent reviewing exceptions the system flags, comes to $28,000. That's the honest total-cost number, not a subscription invoice.

Now the value side. Say the automation saves 18 minutes of staff time per return on the collection and organizing work alone. Across 800 returns, that's 240 hours freed up. Value that time at $140 an hour, a reasonable blended rate for the level of staff who do that chasing and reviewing, and you get $33,600 in value created.

Run the formula: ($33,600 − $28,000) ÷ $28,000 = 0.20, a 20% return in year one. That's not the triple-digit number a vendor's slide would show you, and that's fine. It's a real return, and it held up because it came from a total-cost number instead of a licensing invoice.

Now find the break-even point using these same assumptions: at what return volume does value created stop covering total cost? Set the two sides equal: returns × (18 ÷ 60 hours) × $140 = $28,000, which simplifies to returns × $42 = $28,000, or roughly 667 returns. Below somewhere around 650-700 returns a season, at these same inputs, this same automation build costs more than the time it frees up is worth. That's the number a smaller firm needs before signing anything, and it moves the moment your own minutes-saved, hourly value, or total cost differ from these illustrative inputs. Which is the whole reason to run your own math instead of borrowing someone else's.

When the math genuinely doesn't work for a firm your size

Here's the part most vendor conversations skip, because it's bad for their sales cycle: for a firm well under that break-even volume, a two-partner practice with 200 returns and a light advisory book, say, a full custom automation build for document collection alone usually doesn't clear the bar in year one. The fixed costs of a build (implementation, maintenance) don't shrink much just because the return volume is smaller. That mismatch is the actual trap: a vendor selling a fixed-price seat will tell a firm at almost any size that the math works, because their revenue depends on the seat getting sold, not on your break-even number being real.

The honest answer, when a firm sits below its own threshold, usually comes down to timing or scope. Wait until return volume grows into the number that justifies the build, or narrow the target: compliance deadline and IRS notice tracking tends to carry a lower cost floor than full document collection, because the risk it addresses doesn't scale with volume the same way. Sometimes the honest answer for this season really is "not yet," and treating that as a legitimate result of doing the math beats forcing a build that won't pay back.

One more thing worth being direct about: none of this math is about running the firm with fewer people. It assumes your staff is still there, doing the higher-value work the freed-up hours make room for. If a vendor's ROI pitch is implicitly about cutting the team rather than freeing its time, notice that. It's a different pitch than the one this formula is built to evaluate.

FAQ

How do I calculate AI ROI in accounting if I don't have exact numbers yet?

You don't need exact numbers to start. You need honest, rounded ones. Estimate minutes saved per return or per task from what your own staff already tells you eats their time, price that time at a rate you'd defend to a partner, and total every cost the build will touch, not just the invoice. The formula works fine on careful estimates; it only breaks on invented ones.

Why don't vendors show total cost, only licensing cost?

Because licensing cost is the number that makes their pitch look best, and total cost is the number that invites harder questions. Implementation, maintenance, and the staff time spent reviewing exceptions are real costs no matter who sells you the tool. Leaving them off the slide doesn't make them disappear; it just moves the surprise to after you've signed.

Is a 20% ROI in year one actually good for a small CPA firm?

For a first-year automation build, yes. A modest, honestly-calculated positive return is a far better signal than an eye-catching one you can't verify, because it means the math survived real total-cost accounting instead of marketing rounding. It also tends to improve in later years, since the build cost doesn't repeat but the time savings do.

Does AI ROI math mean my firm needs fewer staff?

No. The math in this article assumes your staff is still there to use the hours it frees up, on advisory work, on handling more return volume, on not adding headcount as the firm grows. If a vendor's ROI pitch is really about cutting people rather than freeing their time, that's a different conversation than the one this formula is built for.

The honest version of this conversation only works with your numbers, not the illustrative ones above. Your return volume, your blended staff rate, and your actual total cost once maintenance and review time are counted are the inputs that tell you whether your firm sits above or below its own break-even line. That's a better use of ten minutes than reading one more vendor case study.