Post 4 of 7

RSM surveyed 827 US and 203 Canadian middle-market executives about AI this July. Three of the numbers are worth putting next to each other.

Eighty-six percent have AI integrated into operations. Thirty-six percent have it fully embedded across core processes. Ninety-seven percent say they are satisfied with the business value it delivers.

Almost everyone has it. Around a third have it where it changes how the business runs. And nearly all of them are pleased.

Set that against the number everyone quotes from the other direction. MIT’s 2025 study of enterprise AI found 95 percent of generative AI pilots produced no measurable return. The methodology has been publicly challenged since, and you should know that before you use it at a board meeting.

Both can be accurate, because they measure different things. Satisfaction is what an executive reports about a tool their team uses. Measurable return is what shows up on the P&L. The distance between those two is the subject of this article.

We have been here before. I saw it with ERP, with CRM, with business intelligence, and the pattern each time was the same: the tool arrived, adoption stalled, staff worked their way back to the familiar method, and three years later somebody canceled the licenses without discussion.

Where the money is coming from

The spending is not tentative. Fifty-eight percent of those executives plan to put $1 million or more into AI this fiscal year, and 84 percent expect to increase AI spending next year.

That money is being moved from somewhere, and this is the part of the survey I would pin to a wall. Among firms increasing AI investment, 43 percent are cutting business intelligence and analytics, 41 percent are cutting cybersecurity, and 40 percent are cutting external consulting and advisory.

Then the same survey asks what is stopping them scaling AI. Data quality, at 34 percent. Security and privacy, at 30 percent. Integrating legacy systems, at 28 percent.

Read those two lists together. Firms are defunding analytics and cybersecurity in order to buy AI, and then naming data quality and security as the two things preventing the AI from working. That is not a technology problem. It is the execution gap with a budget line attached.

I will note the third item on the cuts list without arguing with it. Forty percent are reducing external advice, which includes people like me. What I would say is that the firms cutting their data and security capability to fund an AI program are the ones most likely to need help within eighteen months, and they will be buying it in worse circumstances than these.

The questions worth asking instead

Most of the AI conversation aimed at owners is written by people selling AI. It asks whether you are behind. That framing produces bad decisions, and it is the reason so much money is about to go into pilots that produce nothing.

Top down

The owners I talk to who are getting somewhere ask a narrower question: which two processes in this firm consume the most senior time for the least client value, can a machine do part of that acceptably, and how much human judgment is required.

There is a follow-on question of equal weight. Are those processes solid, outdated, or a workaround nobody ever went back and fixed? Automating a workaround gives you a faster workaround and a longer-term problem.

Those questions have a short list of answers in almost every professional services business.

Proposal and document production. First drafts, prior-work retrieval, formatting, tailoring boilerplate to a client. Senior people spend hours here and clients pay for none of it.

Research and first-pass analysis. Summarizing source material, finding comparable data, assembling the background section that a junior used to build over two days.

Meeting capture and follow-up. Notes, actions, the summary email, the file note that half your team writes late on Thursday and the other half never writes at all.

Internal knowledge retrieval. The question “have we done something like this before” currently answered by asking three people and hoping one of them remembers.

None of that is glamorous. All of it is senior hours going into work nobody bills, in a business where senior hours are your scarcest asset and your largest cost.

Bottom up

The mistake I see most is running this in one direction only.

Ask your team the same question. Where do they see two processes that are good candidates for automation and would give them leverage to do more advanced work? They will name things you have never seen, because you have not done that part of the job in six years. The follow-on question applies to them too: is the process solid, outdated, or a workaround, and are the steps in it still the right steps?

Then ask the question almost nobody asks until it is too late. How does this change roles, responsibilities and the structure of the organization? If a junior analyst’s first two years were built around work a machine now does in an afternoon, your training model just changed, and so did the path you offer someone joining next September. Better to decide that deliberately than to discover it when your best second-year associate resigns because the job stopped being interesting.

Bottom up also decides adoption, and the RSM survey puts a number on why. Eighty-five percent of those executives said leadership enthusiasm for AI exceeds employee enthusiasm. That gap is the adoption problem stated as a statistic. People implement what they helped design. They work around what was handed to them.

The ROI question, answered honestly

Owners ask what a realistic twelve-month return looks like. Vendors quote percentages that assume everything works. What I have seen in practice starts smaller and arrives sooner: five hours a week back for one person, a couple of days a month across a team.

Here is how I would model it. Take one process from the lists above. Count the hours it consumes across the firm in a month, at loaded cost, not billing rate. Assume a tool plus a reworked process removes 30 to 40 percent of those hours in the first year, not the 80 percent in the sales deck, because your people will spend time checking output, some of it will be wrong, and they will need to modify how they work.

Then decide what happens to the hours that are now available, and what new skills the team should build so they can have a stronger impact with them. This part decides whether the investment returns anything. Hours that convert to billable delivery or business development produce a return. Hours that convert to slightly less pressure produce a happier team and no measurable change on the P&L, which may be a legitimate choice and should be made on the record.

Firms that do this arithmetic before buying tend to buy less and use more of what they buy.

Why the embedded number is the one that decides it

The gap between 86 percent using AI and 36 percent having it embedded does not mean two thirds of firms need to hire a data scientist. For a 40-person professional services business that would be a mistake.

What it means is that adoption needs an owner inside the firm, with time allocated, standards to set, and the authority to change how work gets done. It also needs a budget, for internal capacity as much as for the software. Somebody has to decide which tools are permitted, what client information may never go near them, how output gets checked before it reaches a client, and what the standard is for a document a machine helped produce.

Write the reasoning down alongside the rules. A policy that says “no client data in external tools” gets worked around within a month, the first Friday afternoon a deadline stands between somebody and getting home at a reasonable hour. The same policy, with two lines explaining which client obligations it protects and what happens to the firm if one of them is breached, gets followed by people who then apply the same logic to the situation you did not anticipate. Rules cover what you thought of. Reasoning covers the rest.

Without that person, three things happen. Your team uses these tools anyway, on personal accounts, what gets called shadow usage, with client material in it, whatever their intentions were about taking it out first, and you find out later. Quality varies by whoever is at the keyboard. And nothing changes in how delivery is structured, so you pay for tools and keep the old cost base.

On the fear underneath

A version of this question comes up privately with almost every owner, and it is rarely asked in the room with the team present.

If a machine can do a first draft, research, and analysis, what happens to a business that sells expertise by the hour.

My read, for what it is worth. The part of your service that is information retrieval and document production is going to cost less, and clients will expect to pay less for it within a few years. The part that is judgment, accountability, relationship, and being the person a client calls when something has gone wrong is not close to being automated, and it is the part your clients value most when they are under pressure.

The firms at risk are the ones whose value sits mostly in the first category and who price as though it sits in the second. If your fees are largely justified by volume of output, that is worth confronting this year on your own terms.

The firms in a strong position are the ones who use the cost reduction to do more of the judgment work, at better margin, for clients who want a person accountable. That was always the better business. This is going to make the difference between the two more visible.

Where I would start on Monday

One process for your seniors. The one where they lose the most time to work clients do not pay for. Measure the hours it takes now, at loaded cost.

One process for the team. The one they lose the most time to, or the one clients push back on when they see it itemized on the bill. Measure that the same way, and ask whether automating it would produce an advantage over the current approach or just a faster version of something that was never right.

One person named as owner, with hours protected for it, a budget, and a one-page policy on client data that carries the reasoning as well as the rules.

One number to judge it on in ninety days, decided before you spend anything.

Three months of that teaches you more about what this technology does for a firm your size than a year of reading about it. And it keeps you out of the group that spends capital on capability they have nobody to operate.

---

Sources

Figures in this article were taken from the sources below on 14 September 2026. Each entry gives the reference period the data covers and the date it was published, so you can judge how current a number is whenever you happen to read this.

RSM US, Middle Market AI Survey 2026. Survey of 827 US and 203 Canadian middle-market executives, published 21 July 2026. Retrieved 14 September 2026.

https://rsmus.com/insights/services/digital-transformation/rsm-middle-market-ai-survey.html

https://rsmus.com/insights/services/digital-transformation/rsm-middle-market-ai-survey/rise-ai-spend-money-going.html

MIT NANDA, The GenAI Divide: State of AI in Business. Published 2025. The 95 percent figure has been challenged on methodology since publication. Retrieved 14 September 2026.