Walk any accounting technology conference floor and everything sounds wonderful. Booth after booth of genuinely impressive software, each solving something you recognise. You come home with a shortlist, and somewhere in the following quarter a licence gets bought.

Then very little happens.

This is the position most client advisory practices are actually in, and it is not a capability gap. The tools have been acquired. Copilot is deployed across the firm, the ERP vendors ship AI features every quarter, and half the stack contains automation nobody has switched on. What is missing is the machinery that turns a purchase into a change in how the work gets done, and no amount of model quality substitutes for it.

It shows up in how firms measure themselves. Adoption gets reported as licences deployed and seats activated, which are purchasing metrics. Almost nobody reports processes changed, which is the only number that would tell you whether any of it worked.

That claim is testable, and the test does not favour the prevailing view.

What the Benchmark Data Actually Rewards

The CPA.com and AICPA PCPS CAS Benchmark Survey, covering 206 practices and calendar year 2023 data, is the closest thing this service line has to a scoreboard. The headline is healthy: median growth of 17 percent, and net client fees per professional of $156,250, up 29 percent in two years. CAS is the fastest-growing area in public accounting and nobody disputes it.

Look at what separates practices inside that population, though, and almost none of it is a technology decision.

A plan. A niche. Dedicated people. None of those arrive with a subscription.

The niche finding is the most interesting of the three, because it is usually read as a marketing advantage and it is really a delivery one. Serve twenty restaurants and you meet the same handful of point-of-sale systems, the same chart of accounts, the same revenue recognition questions and the same month-end rhythm. The second engagement is cheaper than the first, and the twentieth is cheaper again. Specialisation is how a practice buys repeatability, and repeatability is what makes a technology rollout worth doing at all.

The technology finding deserves care rather than dismissal, because the effect is real: practices investing continuously in technology serve a median of 100 clients against 67 across all respondents. But note the shape. That is a ceiling being raised on a practice that already has structure underneath it. A firm without the structure does not get 100 clients from buying software. It arrives at chaos faster.

The Selection Test Most Firms Skip

Before any of the delivery discussion, there is a distinction that firms routinely fail to make, and it is expensive.

Automation is rules-based, repetitive and predictable. You set it and it does the same thing every time. AI is probabilistic. Prompt it twice and you are extremely likely to get different results, unless the environment has been constrained heavily.

That is not a ranking. It is a selection criterion, and skipping it produces two failure modes that look nothing alike. Buy AI for a problem that wanted a rule, and you get inconsistency where you needed determinism, then conclude AI does not work. Automate something that genuinely needed judgment, and the exception queue grows until somebody quietly turns the automation off.

There is a longer-run argument for holding this line carefully. We are working with the worst AI we will ever have, which is usually offered as a reason for patience. It cuts the other way too. If the models keep improving, the durable investment is not this quarter’s tool. It is the structure that will still be there to absorb the next one.

The Machinery, Written Down

So what does a delivery function consist of? Every firm claims to onboard clients well. Very few can describe how, and the description is the whole difference.

The same discipline governs buying. Before evaluating any system, write requirements: the core things the system must actually do. Then check the candidate against what you already own, because the most common expensive mistake in a firm tech stack is paying twice for the same functionality. That single habit would deflate most conference shortlists, including the impressive ones you find yourself wanting a use case for.

It is also what makes upskilling more than a slogan. Removing compliance work is supposed to free people for advisory conversations, and that only happens if somebody protects the reclaimed hours from being refilled with different low-value work. Absent that, automation returns capacity to a queue rather than to a career.

The Function Nobody Funds

Put the pieces together and the constraint has a name. Most practices have no repeatable way to take a capability, decide whether it needs determinism or judgment, pilot it on one engagement, train people, and roll it out without overwhelming them. That is project management, and firms have historically declined to resource it on the grounds that it does not bill.

Which turns out to be a choice rather than a fact.

At CohnReznick, where Michelle Voyer, CPA, directs the software solutions group, the project management office and the offshore centres of excellence within client advisory services, the project managers run technology implementations as chargeable engagements. Systems design, approval workflows, segregation of duties, integration between third-party tools and the ERP: all of it scoped, tracked and invoiced. A project management office in that model is not overhead wrapped around the real work. It is a service line.

That is the part worth stealing, and it dissolves the objection. Delivery capability does not have to be justified as an investment against future efficiency, argued for annually, and cut when utilisation dips. It can be sold.

The alternative is familiar enough to describe. Someone runs a promising pilot on one engagement, it works, and nothing follows, because scaling it was nobody’s job and the person who ran it went back to chargeable work. The capability was real. The mechanism to spread it did not exist.

If you have one decision to make this year and you already hold an AI licence your teams are not fully using, the return is not in a better model. It is in the person whose job is making sure the model reaches the work.