A study analysing 1.4 million real workplace AI interactions found that just 12 percent of companies using AI tools achieve actual business impact.¹ Not 12 percent seeing transformational results while the rest see modest gains. Twelve percent seeing any measurable return at all. The other 88 percent bought licences, ran pilots, attended webinars, and are now sitting on tools that cost money every month and produce no discernible change to output, speed, or margin.
If you are the principal or operations lead at a professional services firm and this number sounds familiar, this article is for you. Not to sell you another tool. To explain what the 12 percent are doing differently, and why the other 88 percent are stuck.
The three things the 12 percent do
The KPMG and UT Austin research¹ found that the highest-impact AI users do not optimise prompts. They frame problems, guide thinking, and take ownership of outcomes. That distinction is worth unpacking, because it points to three concrete operational choices that separate firms seeing returns from those that are not.
They automate a complete workflow, not a single task.
The most common failure pattern I see is what I think of as task-level automation. Someone uses ChatGPT to draft emails more quickly. Someone else uses it to summarise meeting notes. Each of these things saves a few minutes. None of them change the economics of the firm, because the surrounding workflow is unchanged. The email still needs to be reviewed, approved, sent, tracked, and followed up by a human. The meeting notes still live in a document no one acts on.
The firms seeing impact automate the whole loop. A client onboarding process that previously required five separate manual steps across three tools now runs end-to-end, with a human reviewing the output at a single checkpoint rather than touching every stage. This is the difference between Level 1 (education, using a tool you already have more effectively) and Level 2 (integration, connecting two or three tools so they pass information between them automatically). Most firms dabbling at Level 1 and calling it an AI strategy are not going to see the 12 percent’s results.
They measure output before and after.
This sounds obvious. It almost never happens. I regularly ask clients how long a specific process takes before we touch it. Most cannot tell me with any precision. “It varies” or “not that long” are the usual answers. When we instrument the process properly, the actual number is often startling: a client intake workflow that felt like it took about twenty minutes turns out to involve three hours of fragmented effort spread across two people and four days.
You cannot demonstrate impact from a change you did not measure before you made it. The firms achieving returns picked a specific process, timed it, counted the steps, noted the error rate, and then changed one thing. This is not sophisticated project management. It is the discipline of writing down a before number so you have something to compare the after number to.
They assign one person accountability for the result.
Shared ownership of an AI implementation is no ownership. When a pilot sits across the operations lead, the IT manager, and whoever went to the conference, it belongs to no one. Questions do not get answered. Problems do not get escalated. The tool drifts into the category of things the firm is “exploring.”
The 12 percent assign a named individual responsibility for whether the implementation works. That person owns the metric, owns the timeline, and owns the decision about whether to continue or stop. This is the same governance principle that applies to any other operational function: fraud risk ownership, compliance monitoring, client communication standards. If no one can be asked “how is this going and what do the numbers say,” the answer is usually that it is not going anywhere.
The firms seeing returns from AI did not buy better tools. They made three decisions about how to run the work: automate the full workflow, measure before and after, and put one person’s name against the outcome.
Why the 88 percent stay stuck
The failure modes are not mysterious. They cluster around a few patterns.
The first is adopting AI at the wrong level. A firm that needs Level 2 integration spends six months trying to solve the problem at Level 1, gets frustrated, and concludes that AI “doesn’t work for us.” Level 1 (education) is the right starting point for most questions, but it has a ceiling. If you need two systems to talk to each other automatically, a better prompt will not fix that.
The second is confusing activity with progress. Running a pilot, attending training, adding a tool to the stack: these produce the feeling of forward motion without producing measurable change. The research from KPMG and UT Austin¹ was pointed about this. Optimising prompts, the activity most people associate with “using AI well,” was not what distinguished high-impact users from low-impact ones. Framing the problem correctly and owning the outcome were.
The third is governance that follows tools rather than actions. A useful reframe, one that I have seen work in regulated professional services environments, is to stop asking “what do we do with this new tool?” and start asking “what business actions do we want to govern, and who owns each one?” The tool is just the mechanism. The action, and the accountability for it, is what matters. This is particularly relevant in FCA-regulated firms, where the direction of travel from the regulator is clearly towards named accountability for AI-assisted processes.
What to do with this
None of the above requires a large budget or an engineering team. Level 2 integration work typically takes days to weeks and costs a few hundred to a few thousand pounds, depending on the complexity. The prerequisite is not money. It is the willingness to pick one specific workflow, instrument it before you touch it, change the complete loop rather than one step, and assign someone ownership of whether it works.
If you are a principal at a professional services firm who has bought tools and is not seeing returns, the question is almost never “which tool should I add?” It is nearly always “which workflow should I close the loop on, and do I have a number for what it costs today?”
If working through that question with someone who has done it in firms like yours sounds useful, a discovery call is a reasonable place to start.
Sources:
[¹] KPMG and University of Texas at Austin joint research analyzing 1.4 million workplace AI interactions. Available at: kpmg.com/us/sophisticated