The Clarity Problem
Why AI investment keeps failing to deliver and what organizations can do about it
Goldman Sachs made headlines earlier this year when chief economist Jan Hatzius declared that AI had contributed “basically zero” to US economic growth in 2025. It’s a striking number, but it largely reflects where AI hardware gets manufactured, not whether AI is working inside organizations. The more relevant finding comes from a survey of nearly 6,000 executives: 70% of firms are actively using AI, and 80% report no measurable impact on productivity.
That’s the gap worth examining.
These numbers don’t suggest a technology problem. They suggest is a problem in how organizations are thinking about AI: the questions they’re asking, the understanding they’ve built, and whether any of it connects to the actual work they’re trying to do.
As I see it, there are four specific drivers behind the gap.
01 — People have mistaken AI marketing for AI literacy
The information environment around AI is not primarily educational. It’s promotional. The same mechanics that make media exhausting are operating here: fear of falling behind drives engagement the way outrage drives clicks. The result is a workforce that has absorbed enormous amounts of AI content without developing real AI understanding. People know the vocabulary and have heard the promises, but most lack a working model of how these systems actually function: what their failure modes are, where they genuinely excel versus where they confidently underperform, or how to evaluate a vendor claim critically.
That anxiety is not an accident. It’s the business model. Organizations need to develop real AI fluency as a way to counter it.
02 — Fragmented, individual experiments don’t yield organizational gains
The dominant model for AI adoption has been a kind of distributed volunteerism: individuals and teams discovering tools, experimenting in isolation, achieving small wins or not, and moving on. Some organizations have accelerated this by encouraging everyone to “try AI” — an approach that produces enthusiasm, anecdotes, and very little compounding value.
Individual productivity gains don’t automatically translate to organizational capability. According to Section’s three-year study of 5,000 enterprise knowledge workers, 69% remain in experimenter mode while fewer than 4% have reached proficiency or expert level. The result is an organization with scattered activity and no shared foundation to build on: experiments don’t connect, learnings don’t accumulate, and each team solves a slightly different version of the same problem in isolation.
The missing ingredient isn’t more experimentation. It’s collective intelligence and strategic direction.
03 — AI is being treated as a solutions problem, not a systems one
When organizations evaluate AI primarily on performance (speed, accuracy, cost savings), they’re asking a solution question about something that is fundamentally a systems issue. Inserting AI into an organization doesn’t happen in isolation. It lands inside an existing ecosystem that will have to adapt: workflows change, roles shift, training requirements emerge, performance standards need rethinking, and sometimes the products and services themselves need to evolve to reflect what’s now possible.
All of that costs time, attention, and organizational energy, and the question that almost never gets asked explicitly is whether the return justifies it. The assumption embedded in most AI adoption conversations is that the answer is always yes. It isn’t.
The organizations that treat AI as a discrete tool rather than a systems intervention are at risk of underestimating the scope of adaptation required and overestimating its potential gains.
04 — We’re asking the wrong questions
The quality of any solution is almost entirely determined by the quality of the problem statement defined up front. Frame it wrong at the start and everything downstream suffers — not because of bad decisions along the way, but because you were solving for the wrong thing from the beginning.
The questions driving most AI adoption are efficiency questions: How do we save money? How do we produce more? How do we move faster? How do we reduce headcount? These are not unreasonable things to want, but they are not value questions, and the difference matters enormously.
Efficiency questions treat AI as a cost-reduction mechanism, where value questions ask what becomes possible that wasn’t before — what AI allows people to do better rather than just faster, and whether the work being accelerated is the right work in the first place.
A better starting point
Most of the AI conversation is working against clarity because clarity doesn’t drive clicks the way urgency does. But clarity is what’s needed, and it comes not from the feed but from honest questions about the actual work and what you’re trying to accomplish. Before reaching for a tool, the more useful questions sound like this:
Quality: What do we want to produce that we can’t produce well enough today?
Speed: Where are we spending time on work that doesn’t require human judgment?
Capability: What have we never been able to do, and does that still have to be true?
Process: Where does friction live in our current workflows?
Direction: What has changed with our customers or the market that requires us to make a change?
These aren’t prompts for a brainstorm. They’re meant as a frame for evaluating whatever crosses your feed against something you’ve identified as worth solving. That’s a different relationship to the information environment, and it doesn’t require disengaging from AI, but learning a healthier way to consume and respond to it.
This isn’t new, but some of the conditions are.
Every major technology shift has produced a gap between early investment and realized value: a period of hype and expensive lessons before organizations figured out how to integrate the new capability in ways that actually worked. The difference this time is speed, scale, and opacity.
Previous transitions played out over decades, with enough time and friction for course correction to happen organically. This one is moving faster than most organizations can assess it, with investment levels that create enormous pressure to show returns before the foundations are in place. And unlike most prior technologies, the systems at the center of it are genuinely difficult to audit — black-box models, opaque company practices, and no reliable framework yet for measuring what AI use actually produces at the organizational or economic level.
The organizations that will navigate AI well are not the ones that find the right tool. They’re the ones that develop shared understanding, internal judgment, and the decision-making frameworks to evaluate tools critically, adapt when circumstances change, and build on what they learn.
That’s not a technology investment. It’s a people and systems one, and it comes before deployment, not after.