KPMG recently found that nearly half of executives who deployed AI agents ended up pulling back once the costs outpaced the benefits. Forbes covered the research in early August. If you run an accounting or bookkeeping firm and you’ve been watching the AI agent hype from the sidelines, that number should tell you something useful: the problem isn’t the technology. It’s how people are buying and deploying it.

Most firms that pull back didn’t fail because agents can’t do the work. They failed because nobody calculated the cost-per-task before they signed the contract. They bought a platform, pointed it at a broad problem like “automate our tax prep” or “handle client questions,” and watched the compute bill climb while the actual hours saved stayed murky. Six months later, someone in finance asks what this thing is actually worth, and nobody has a clean answer.

For a firm doing $1M to $25M in revenue, that’s not an abstract cautionary tale. It’s a decision you’re probably facing right now, and the KPMG data gives you a much better way to make it.

The real question isn’t “should we use AI agents”

It’s “what does this specific task cost us today, and what would it cost with an agent doing it.” That distinction matters more than any vendor pitch you’ll hear this year.

Firms that pull back on AI usually made a category-level bet. They rolled out a general-purpose assistant across the whole practice and hoped efficiency would show up somewhere. Firms that stick with AI, and actually see margin improvement, made task-level bets. They picked one recurring, well-defined piece of work, measured what it cost in partner and staff hours before automation, then measured the same thing after.

This is the same discipline good bookkeepers apply to a client’s P&L. You don’t guess at a variance. You isolate it, quantify it, and decide if it’s worth fixing. Apply that lens to your own AI spend and the pullback risk mostly disappears.

Where the hours actually go in a $1M-$25M firm

Before you can calculate cost-per-task, you need to know where the tasks live. In firms this size, we consistently see three concentration points.

Month-end and year-end crunch. A huge share of firm capacity, often somewhere in the 30-50% range of total staff time, gets consumed in a four-week window around close. Bank feeds need reconciling, AP and AR need tying out, payroll entries need checking, and partners need a close pack they can actually review instead of a raw export. This work is repetitive, rules-based, and painfully manual, which makes it exactly the kind of task where cost-per-task math works in your favor.

Client onboarding drag. New clients hand over a shoebox of PDFs, three different accounting systems’ worth of history, and an expectation that everything will be clean within weeks. Document collection, chart-of-accounts setup, and historical clean-up routinely stretch onboarding into a quarter or longer. Somewhere between 20% and 30% of new clients delay billable work that long, which means you’re carrying acquisition cost with no revenue to offset it.

Advisory time crowded out. Compliance work fills the calendar first because it has hard deadlines. Advisory conversations, the ones billed at 2-3x the compliance rate, get pushed to “next month” indefinitely. This is the quiet cost nobody puts in a KPI deck, but it’s usually the largest one.

Each of these is a candidate for an agent. None of them is a candidate for a vague, firm-wide AI rollout that you can’t measure.

What cost-disciplined agent deployment actually looks like

Here’s the difference between the deployments that get pulled back and the ones that stick, told through three specific agents we build for firms like yours.

The Month-End Close Agent pulls bank, AP, AR, and payroll feeds automatically, reconciles them against the ledger, flags variances that fall outside normal range, drafts the journal entries, and assembles a partner-ready close pack. Before you deploy anything, you can measure exactly what close currently costs: X staff hours at Y loaded rate, across Z clients, concentrated in that four-week window. After deployment, you measure the same thing again. There’s no ambiguity about whether it’s working, because you defined the task narrowly enough to track it.

The Client Onboarding Agent runs a guided workflow that collects documents from new clients, builds out the chart of accounts, and produces a clean opening trial balance without a partner chasing down a client for the fourth time about a missing bank statement. Again, the cost-per-task is knowable up front. You already know your average onboarding cost in hours. You can price the agent against that number rather than against a promise.

The Advisory Insights Agent reads each client’s monthly numbers, surfaces the three things worth discussing, and drafts talking points before the partner meeting. This one is a little different because its value shows up as revenue capture rather than cost avoidance. It’s easiest to justify by tracking a simpler number: how many additional advisory conversations happen per partner per month, and what those conversations bill at. If that number doesn’t move, you’ll know within a quarter and you can adjust or walk away, which is exactly the discipline that prevents the KPMG pullback pattern.

You’ll notice none of these are “an AI agent for the firm.” Each targets a bounded, measurable task with a known before-cost. That’s the entire difference between the 49% who pulled back and the firms quietly compounding margin gains.

If you want the fuller mechanics of how these agents connect to your existing systems without a rebuild, our ops automation work walks through the integration side, and our advisory tooling covers how the insights layer feeds partner conversations specifically.

The dollar reality for your firm

We typically see firms this size leaking somewhere in the range of $60,000 to $180,000 a year across these three pain points combined. That’s not one dramatic failure. It’s the accumulation of month-end overtime, delayed onboarding revenue, and advisory hours that simply never got booked because nobody had the calendar space.

Here’s a rough way to check your own number. Take your average loaded cost per staff hour. Multiply it by the extra hours your team pulls during close season versus a normal month. Add the billable revenue you’re not collecting because onboarding drags a quarter for a fifth of your new clients. Add the advisory hours you know you should be billing at 2-3x compliance rates but aren’t, because the calendar never has room. Most partners who run this exercise for the first time are surprised by how quickly it adds up, and more than a few land right in that $60K-$180K band.

That’s the number you should be comparing against any AI agent’s price tag, not the vendor’s marketing deck.

Why the KPMG pullback happened, and how to avoid repeating it

Digging into the pattern behind KPMG’s finding, the common thread among executives who pulled back wasn’t that agents underperformed technically. It’s that nobody had defined success before deployment. Costs crept up through usage-based pricing, integration overhead, or maintenance that wasn’t budgeted, and when someone finally asked “is this worth it,” there was no baseline to compare against.

Firms in our network avoid this by doing three things before they sign anything:

First, they pick a single task with a clear, current cost. Not “improve efficiency across the practice.” Something specific like “the 40 hours per month our senior bookkeeper spends reconciling multi-entity clients.”

Second, they set a review point, usually 60 to 90 days out, where they compare actual hours saved against the agent’s running cost. If the math doesn’t hold, they stop or renegotiate. No sunk-cost momentum.

Third, they treat the rollout as reversible. An agent that handles month-end close should be swappable or adjustable without touching your whole tech stack. That flexibility is what protects you if the cost curve shifts.

This is the same posture we recommend before any Omni deployment, and it’s the reason we start every engagement with numbers rather than a pitch.

A structured way to check before you commit

If you’d rather see this mapped against your own close process than do the math cold, we put together the Month-End AI Close Map for Accounting Firms, a practical worksheet that walks through where hours concentrate during close and where an agent can realistically absorb them. It’s built for exactly the cost-per-task exercise described above, so you’re working from your own numbers rather than a hypothetical. You can grab the close map here and run it against your last close cycle before you talk to anyone about deployment.

That’s a good first step on your own. But the more reliable way to get accurate cost-per-task numbers, without guessing, is to have someone map your actual workflow against real agent costs before you commit to anything.

What the Omni Audit actually does

This is the point where most vendors would tell you to buy software. We’d rather show you the math first.

An Omni Audit is 60 minutes, no deck, and no sales pitch disguised as a “consultation.” We walk through your actual month-end process, your onboarding workflow, and your advisory calendar, and we come back with three things: a clear picture of where your hours are going right now, a cost-per-task estimate for the specific work an agent could take on, and a straight answer on whether the math works for your firm size and client mix. If it doesn’t work, we’ll tell you that too. That’s the whole point of measuring before you deploy.

If you want to see how this applies specifically to firms your size, see Omni for accounting and bookkeeping before booking anything. It’s a good way to understand the scope of what gets audited and what doesn’t.

You can also read more on how other firms in the space are approaching this in our insights library, or browse the broader guides section if you want more background on how agent-based workflows differ from the chatbot-style tools most firms tried first.

The move that actually protects your margin

The lesson from KPMG’s data isn’t “don’t use AI agents.” It’s “don’t deploy AI agents the way you’d buy new practice management software.” Cost-per-task discipline is what separates firms compounding real margin gains from the 49% quietly walking back their investment a year later.

If you’re carrying $60K to $180K of leakage across close season, onboarding drag, and crowded-out advisory time, the fix isn’t a broad AI initiative. It’s three or four narrow, measured deployments, each judged against its own number.

The fastest way to find your number is to look at it with someone who’s done this math for firms like yours before. Book a 60-min Omni Audit and we’ll walk through your close process, your onboarding pipeline, and your advisory calendar together. Or start with the AI audit for accounting and bookkeeping to see exactly what we look at first.

Either way, run the numbers before you sign anything. That’s the one habit that keeps you out of the half that pulls back. If you want to compare notes with other firm owners working through the same decision, our blog has more detail on how the Month-End Close Agent and Advisory Insights Agent get built out client by client, and you can book my Omni Audit whenever you’re ready to see your own figures instead of an industry average.