56% of CEOs are getting nothing from AI. Here's the part nobody's talking about.
PwC just published its 29th Global CEO Survey. 4,454 CEOs. 95 countries.
The headline finding: 56% of companies are getting nothing from AI. No revenue gains. No cost savings. Nothing. Only 1 in 8 have achieved both.
PwC's explanation: companies are running "isolated, tactical AI projects" that don't deliver measurable value.
That's accurate. But it's describing what's happening, not why.
I've been working on the same question for 200+ products over 15 years. Here's what I keep finding underneath every AI deployment that stalls.
It's not the model. It's not the data quality. It's not the governance framework.
It's that the user you designed for doesn't exist.
The Most Common AI Story Right Now
Every enterprise AI project I've walked into has a user persona.
Usually a document. Sometimes a poster on a conference room wall. Always optimistic.
This person has time. Two free hands. Full battery. They read tooltips. They're engaged and curious. They look a lot like whoever facilitated the research sessions.
They do not exist in production.
The real user is tired. Running behind. Three other tabs open. Message from their manager sitting unanswered. They know what they need to do and have about 90 seconds before the next thing pulls them away. Not curious about your interface. Trying to get through their shift.
That gap between the persona and the actual person is where your adoption numbers live.
Four Projects. Same Root Cause.
Clinical Decision Support: 11% → 68% physician adoption
The original interface was thoughtful. Showed AI reasoning clearly. Surfaced recommendations logically. A capable team built it well.
11% adoption.
The assumption baked into every design decision: physicians reviewing AI recommendations are in a state of focused, deliberate evaluation.
They're not. Hour 11. 14-hour shift. Seven other alerts open. Thinking about the patient they just left and the one they haven't seen yet.
We redesigned for that person, not the persona. Trust architecture built for 90 seconds, not 9 minutes. Recommendations that surface differently under time pressure.
68% adoption in 4 months. Same AI. Same hospital. Same physicians who supposedly "hate new systems."
Caregiver Platform: 12% → 68% engagement
The product was designed for patients. Engaged, health-literate patients actively managing their own care. The research confirmed them. The UX served them well.
12% engagement.
We watched who was actually using it. Not patients. Family members. A spouse, an adult child managing a parent's care while running their own life. In a hospital parking lot. One hand free. Anxious. Half battery.
Not a patient. A caregiver. Invisible in the original brief.
68% engagement after we redesigned for who actually showed up.
Agentic EMR: 78 minutes → 8 minutes per anesthesia case
The original workflow: 7 systems, manual data transfer, 78 minutes per case average.
Every system designed for someone at a desk with two free hands and time.
The real user: perioperative environment, patient on the table, one hand free at best. Cannot context-switch between 7 systems without clinical risk.
We didn't optimize the UX. We accepted the physical reality of where documentation actually happens.
Voice-first. Single command. Updates clinical records, pharmacy, billing, and inventory simultaneously.
8 minutes per case. Every case.
340B Compliance Intelligence: built for the wrong expert
The existing tools were designed for the compliance architect, someone who understands the full 340B eligibility framework at a structural level.
The actual user at the point of decision: a frontline pharmacist asking one question. "Is this prescription covered?"
Not the architecture question. The transactional question. One answer, with enough reasoning to defend it if anyone asks.
The whole product was built for the compliance expert. The pharmacist got a firehose of information when she needed a yes, a no, and a sentence.
How we work: Surface, Architect, Prove
After enough projects where we found this gap too late, we changed how we start every engagement.
Three moves. Before wireframes. Before research. Before anything.
PHASE 01: SURFACE · Pattern Recognition
What teams feel: "Users aren't adopting." The brief that reflects stakeholder consensus, not ground truth. The problem as imagined.
What we surface: The real problem named. Org dysfunction mapped. Workarounds documented. "Success" redefined for production conditions.
PHASE 02: ARCHITECT · Defensible Design
What briefs ask for: Intuitive interfaces. Ideal users. Demo-polished. Stakeholder-approved. Optimized for the room, not the ward.
What we build for: Tired users. High stakes. Accountability chains. Interfaces people can explain to a supervisor, an audit, a review board.
Not intuitive. Defensible.
PHASE 03: PROVE · Adoption Metrics
What teams call success: CSAT scores. Controlled test rates. Launch-day numbers. Training sign-offs. Metrics from the room.
What we measure: Production adoption. Workflow integration. Before/after in business terms. The gap between demo and reality, closed.
Brief → Surface → Architect → Prove → Adoption.
About the author
Founder, Thirty Two North. 15 years enterprise UX. 200+ complex B2B products across healthcare IT, fintech, logistics, and enterprise SaaS. 6 iF Design Awards. Pattern recognition is the service.
Follow on LinkedInJoin the conversation
This article was originally published on LinkedIn. Share your thoughts, challenge assumptions, or connect with the author there.
Comment on LinkedInMore from this series
Edition 01: The AI Trust Paradox
Why users reject AI interfaces they'd otherwise trust. Kshitij Kaushik on trust architecture for enterprise AI adoption.
Edition 02: Reasoning Transparency
What 'show your work' actually means in enterprise AI. Progressive disclosure for healthcare IT, fintech, and enterprise SaaS.
Edition 03: The Override Problem
Why giving clinicians an override button doesn't solve AI trust. Designing for human-AI disagreement in enterprise software.