"AI is a tool looking for a process." — Chief Financial Officer (CFO), mid-market manufacturer
We are starting to have conversations with clients who have spent a lot of time and money on AI projects without a clear understanding of whether they are getting a financial return on their investment. However, the first step is to understand that the financial return of Artificial Intelligence (AI) depends far less on the model or toolset you choose and more on the processes you choose to improve with it.
Most AI deployments are at the desktop or departmental process level, or inside a single departmental silo, and the processes they improve were never designed to convert output into dollars. As a result, some firms are beginning to realize that they are only using AI to augment busy work.
That framing isn't mine. A CFO client shared it recently, and he wasn't being dismissive. He had already watched AI do real work in his business. What he wanted was narrower and harder to get. He wanted proof that AI was moving the financial needle and clarity about where else to apply it for similar returns. Then he listed a handful of high-dollar problems, the kind where a fix could create real returns that show up in the financials. Problems he suspected AI could help with but had yet to see it happen.
I've been chewing on that line ever since, because it explains the current state of AI in the middle market better than most of the research published on the subject.
Read Also: Everyone Is Using AI. Almost Nobody Can Prove It’s Working.
RSM released its Middle Market AI Survey 2026 on July 21, and it is the cleanest read available on this segment to date. The firm surveyed 1,030 senior leaders across the United States and Canada in March 2026, at organizations between $30 million and $10 billion in revenue, with a margin of error of plus or minus 3.1 percentage points. These are not billion-dollar enterprises with dedicated transformation offices. These are companies that look like my clients.
The headline numbers read like a victory lap. 86% have integrated AI into operations. 97% report satisfaction with what it is delivering. 54% say the investment has already exceeded their return expectations, and 58% plan to put a million dollars or more into it this fiscal year.
Then you reach the second layer.
Only 36% have AI embedded across core processes. Asked about their approach, 45% said they prioritize deploying AI where it delivers clear value today, against 17% pursuing transformation across the enterprise.
Nearly every organization in the survey measures return on investment for AI: 99% of them. What they measure is process efficiency (40%), productivity and time savings (38%), and decision quality or speed (36%). These firms are measuring AI for what it already does well rather than for what it must do next, which leaves revenue growth, competitive differentiation, and innovation outside the frame entirely.
Read those together, and the satisfaction figure changes meaning. These organizations are not measuring enterprise impact. They are measuring whether a department got faster.
Firms are optimizing for what is easy to prove rather than what scaling actually requires and declaring success on isolated wins.
A separate 2026 survey by Netrio, conducted by Censuswide, surveyed 401 United States information technology leaders at companies with 200 to 5,000 employees. Their review put a sharper edge on the same finding. 82% have AI in production somewhere in the organization. 26% have it scaled and governed enterprise-wide. That being said, Netrio is a managed service provider with a commercial interest in that answer, so one must consider the source. Even so, a 56-point gap is difficult to explain away.
Boston Consulting Group (BCG) put a number on the cause in The CFO's AI Agenda: From Automation to Advantage, drawing on its AI Radar 2026 research. The firm notes that about 10% of AI success traces to the models themselves and another 20% to the underlying technology platform. The remaining 70% sit in organization, workforce, and skills: the data foundations, process design, and human capability that decide whether a capable model produces a defensible result.
BCG's lens is the finance function specifically, though it reaches the same conclusion as the RSM study, which spans industries: the constraint is the operating model, not the technology. Said more simply, it is not the tool; it is the process one chooses to augment with it.
That 70/20/10 split is the quantified version of my client's remark.
When AI underperforms, the reflex is to question model choice. The failure almost always happened before anyone picked a model. As BCG describes the pattern, a general ledger of raw account codes gives a model nothing to reason across, and automating a fragmented process just scales the fragmentation at speed. The tool worked. The process it improved was never designed to turn that work into financial impact.
Here is what I see in the field, and it is more specific than the research usually gets.
Almost every mid-market AI implementation I encounter sits at the desktop or the tool level. An individual is drafting faster. A team has a copilot inside the application they already use. A department bought a point solution that does one job well. Very little of it operates at the business process level, and almost none of it operates at the business level.
On a spectrum from desktop augmentation to business transformation, the current lean is toward the desktop.
The foundation underneath is not being treated as infrastructure. Data architecture, identity, integration, records governance, and process standardization are firm-wide assets, and they are being handled as project-by-project problems; solved locally if they get solved at all, then left behind when the pilot ends. Nobody owns them as shared plumbing, so every new initiative rebuilds a private version of the same foundation. The survey data noted above tracks this precisely. Data quality is the top inhibitor to AI deployment at 34%, with legacy systems integration at 28%, and among the roughly half of respondents whose pilots delivered only moderate or limited success, data quality (53%) and integration challenges (47%) were the leading reasons.
The consequence is the part executives should be watching, and it runs opposite to the direction they assume. AI is being applied to improve work within silos rather than across the business. Every one of those local wins hardens the silo it sits in. The department gets faster at its own version of the process, using its own definitions, on its own data, and the case for standardizing any of it gets weaker with each efficiency gain. Fragmentation that used to be a manual inconvenience is becoming automated, defended, and measured. The organization is not just failing to fix the 70%. It is investing in the conditions that make the 70% harder to fix later.
That is a strategic decision with real financial consequences, and in most organizations, nobody made it. It happened by default: one tool purchased at a time, one AI power user at a time.
Back to my client's list. He was right to press for proof instead of assuming it. For every problem on that list, the answer sits in the same place: whether the process underneath is coherent enough for AI to move it. Which capability he picks matters far less.
Take reconciliation or cash application. High volume, rule-governed, transactional, and usually consistent already, which is why BCG reports touchless automation rates above 90% in those domains. The same holds outside finance wherever the work is standardized before the AI arrives: service desk triage, claims intake, inspection routing, first-pass contract review. The tool found a process waiting for it.
Now take forecasting or margin analysis across regions. The constraint there is semantic. If "margin" carries three different definitions across the business, no model on the market produces a trustworthy cross-regional answer, and the return stalls, not because the AI failed, but because the organization never agreed on its own vocabulary.
Here’s another example. Ask five leaders what counts as an active customer, a qualified lead, a completed job, or an open risk, and count the answers. Every one of those collisions is a ceiling on what AI can do across the business, no matter how well it performs inside any single department.
So, the honest advisory answer to "will this have a financial impact?" is a question in return. Is the process ready to convert the tool's output into a decision, and that decision into a dollar? Where the answer is yes, impact tends to show up fast. Where it's no, the pilot impresses in the demo and disappears before it reaches the Profit and Loss (P&L) statement, which is the trajectory BCG says the vast majority of pilots follow.
The cost of skipping this diagnosis isn't a failed pilot. It's two or three budget cycles spent proving nothing, a set of silos that got more expensive to unwind, and an executive team that has lost the appetite to fund the work that would have actually paid.
My client was right, and I take the line as a compliment to the technology rather than a critique of it. A tool looking for a process is a genuinely capable tool waiting on the conditions that let it pay off.
The organizations pulling measurable value out of AI are not the ones running the most advanced models. They are the ones that did the boring work first: treated data and process as firm-wide infrastructure, standardized before they automated, agreed on definitions, and built fluency and governance. So, when AI arrived, it had a business process to improve, not just a desktop task to speed up.
Someone must own the question of which processes are worth improving. Unfortunately, in most organizations at this present time, no one does.
GBQ's Business Technology Solutions practice empowers the growth of our clients across six disciplines: risk management, cybersecurity, IT governance, AI and automation, data and analytics, and business systems.
If this article raised questions that you cannot yet answer about your own AI strategy, the first place to start is with a conversation. GBQ’s AI advisory services allow us to assess where your organization stands today, prioritize use cases worth your investment, and help build the governance that lets you prove what AI is returning.
To continue the conversation, schedule time with Doug Davidson, director of GBQ’s Business Technology Solutions practice. Or, contact him directly at ddavidson@gbq.com.
Net Effect is a biweekly column written by Doug Davidson, director of the firm's Business Technology Solutions, published in the firm's Bottomline newsletter. Email ddavidson@gbq.com to have your technology questions addressed in a future column.