A study of 1.4 million real workplace AI interactions found that only 12 percent of companies using AI tools achieve real business impact.¹ The rest bought licences, ran pilots and sat through webinars, and are now paying every month for tools that make no visible difference to output, speed or margin.

If you’re the principal or operations lead at a professional services firm and that sounds familiar, the useful question is what the 12 percent do differently.

Three decisions the 12 percent make

The KPMG and UT Austin research¹ found that the highest-impact AI users don’t spend their effort polishing prompts. They frame problems, guide the thinking and take ownership of outcomes. In a firm, that shows up as three operational choices.

They automate a whole workflow, not a single task.

The most common pattern is task-level automation. Someone uses ChatGPT to draft emails faster. Someone else uses it to summarise meeting notes. Each saves a few minutes, and none of them changes the firm’s economics, because the workflow around them hasn’t changed. The email still has to be reviewed, approved, sent, tracked and followed up by a person. The meeting notes still sit in a document nobody acts on.

Firms that see impact automate the whole loop. An onboarding process that used to need five manual steps across three tools runs end to end, with a person reviewing the output at one checkpoint instead of touching every stage. That’s the difference between Level 1 (education: using tools you already have more effectively) and Level 2 (integration: connecting two or three tools so information moves between them on its own). A firm working only at Level 1 and calling it an AI strategy is unlikely to see the 12 percent’s results.

They measure before and after.

It sounds obvious and rarely happens. Ask how long a specific process takes before anyone changes it and the usual answers are “it varies” or “not that long”. You can’t show the impact of a change you didn’t measure first. Firms that get a return pick one process, time it, count the steps, note the error rate, and then change one thing. None of that is sophisticated project management. The discipline is writing down a before number so there’s something to compare the after number with.

They make one person accountable 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 nobody. Questions go unanswered, problems don’t get escalated, and the tool drifts into the category of things the firm is “exploring”.

The 12 percent give a named person responsibility for whether it works: the metric, the timeline, and the decision to continue or stop. It’s the same principle as any other operational function, whether fraud risk, compliance monitoring or client communication standards. If nobody can be asked “how’s this going and what do the numbers say?”, it usually isn’t going anywhere.

The firms seeing returns from AI didn’t buy better tools. They decided how to run the work: automate the whole workflow, measure before and after, and put one person’s name against the outcome.

Why the other 88 percent stay stuck

The failures aren’t mysterious. They fall into a few patterns.

The first is working 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 decides AI “doesn’t work for us”. Level 1 is the right starting point for most questions, but it has a ceiling. If two systems need to talk to each other automatically, a better prompt won’t do it.

The second is mistaking activity for progress. Running a pilot, attending training and adding a tool all feel like movement without producing measurable change. The KPMG and UT Austin research¹ was pointed about this: prompt optimisation, the thing most people associate with “using AI well”, wasn’t what set high-impact users apart. Framing the problem properly and owning the outcome were.

The third is governance that follows tools rather than actions. A better question than “what do we do with this new tool?” is “which business actions do we want to govern, and who owns each one?” The tool is only the mechanism. In FCA-regulated firms this matters more, because the regulator is clearly heading towards named accountability for AI-assisted processes.

My read

None of this needs a large budget or an engineering team. What it needs is a willingness to pick one workflow, measure it before touching it, change the whole loop rather than one step, and give someone ownership of whether it works. I think the 12 percent figure says less about AI than about how firms run projects. For a principal who has bought tools and isn’t seeing a return, the missing piece is almost never another tool. What’s missing is a workflow with a before number and a name next to it.

Sources:

[¹] KPMG and University of Texas at Austin joint research analyzing 1.4 million workplace AI interactions. Available at: kpmg.com/us/sophisticated