Two numbers from this year's enterprise AI surveys tell the whole story. According to a WRITER survey of 2,400 executives and employees, 59% of organizations now spend at least $1 million a year on AI. Only 29% say they're seeing significant returns. KPMG's Global AI Pulse for Q2 2026 is even bleaker: just 7% of organizations report measurable ROI, and perceived productivity gains actually fell during the quarter.
The money is real. The usage is real. The returns aren't. Something is broken in the middle, and it isn't the model.
the usage data points at the gap
This week, researchers from OpenAI, Columbia, and Wharton posted a working paper analyzing 17 million real ChatGPT Enterprise messages across more than 1,700 organizations. Two findings stand out. First, the heaviest users inside adopting firms are early-career workers — juniors send eight to nine more messages a week than the average active user. Second, in the authors' own words, firms "are still actively learning how to integrate AI into organizational workflows."
Read those together and a picture emerges: AI adoption is happening bottom-up, one individual at a time, as personal productivity. A junior analyst drafts faster. A coordinator summarizes meetings. Each person is a little quicker inside a process that hasn't changed at all. The handoffs, the approvals, the rework loops, the waiting — all still there. That's why the ROI never shows up at the organizational level. You can't measure a process improvement when no process was improved.
Individual productivity gains don't compound. Process redesign does.
this is a mapping problem, and UX people are mappers
Fixing this doesn't start with a bigger model or another platform. It starts with someone drawing an honest picture of how work actually flows — not the official version in the process doc, but the real one, with the workarounds and the spreadsheet nobody admits to. That is a research and mapping exercise, and it's exactly what UX professionals have been trained to do for years. We just called it different things: service blueprinting, journey mapping, contextual inquiry, task analysis.
Blueprint before you buy. Map the end-to-end process — frontstage and backstage — and mark where time actually goes. In most business processes, the expensive part isn't any single task. It's the seams: handoffs, status-chasing, re-keying data between systems. Those seams are where AI belongs, and you can't see them without the map.
Interview the juniors. The arXiv paper tells you where the knowledge lives: the most intensive AI users in your company are the most junior people in it. They've already discovered, through daily trial and error, what the tools are good at within your context. Treat them as research participants. Their improvised workflows are prototypes of your future process.
Redesign the flow, not the task. If AI drafts the report in two minutes but the report still waits four days for approval, you've optimized 2% of the cycle time. The UX move is to question the flow itself: does this approval need to exist? Can AI pre-check the routine cases so humans only review exceptions? That's where the compounding returns live.
the pitch to make this quarter
If you're a UX professional, this is the moment to widen your remit. Nearly a quarter of organizations told KPMG their investors are now demanding proof of business value from AI. Leadership is holding an expensive capability and no map for where to apply it. You have the mapping skills and — per this week's data — a company full of junior colleagues who've already run the field experiments.
Pick one process that everyone complains about. Blueprint it in a week. Mark the three seams where AI would remove real waiting or rework, and instrument them so the before-and-after is measurable. That's not a moonshot; it's a service design project with a new material.
The companies seeing returns from AI aren't the ones with the best models. They're the ones that redesigned the work. Somebody has to be the person who understands the work well enough to redesign it. It might as well be you.