The Experiment and What It Found
A paper published by researchers from Harvard Business School and Stanford tested a question every executive has been implicitly betting on: can GenAI enable people without deep domain expertise to perform at the level of specialists?
They recruited three groups — domain insiders, adjacent professionals, and distant outsiders — and gave them the same task with the same AI tools: conceptualize and then execute a piece of web content. The experiment was designed to isolate one variable. Not model quality. Not prompt skill. Knowledge distance.
For conceptualization — structuring, outlining, ideating — AI equalized performance completely. Every group, regardless of background, performed at specialist level. The gap disappeared.
For execution, it split. Adjacent professionals matched the specialists. Distant outsiders consistently underperformed — and in many cases, made the output measurably worse than if they had worked without AI at all.
The researchers called this the GenAI wall effect: a threshold, defined not by the technology but by the human's distance from the execution domain, beyond which AI cannot bridge the gap.
"GenAI suggested some catchy hooks… Actually, I didn't fully understand what it was doing because I never wrote an article like that. I added random stuff to make it more 'marketing.'"
The mechanism is precise. The participant couldn't evaluate the AI's output because he had no foundation for judging what good looked like. So he removed what he didn't recognize as valuable — the hooks, the narrative architecture, the conversion logic — and replaced them with what felt right to him. He degraded a correct answer because he couldn't see that it was correct.
The AI was fine. The knowledge distance was the problem.
Four Threads. One Finding.
What makes this experiment significant isn't that it's surprising. It's that it's the fourth independent thread — from different disciplines, different levels of analysis — to arrive at the same place.
Naval names the asset. Ismail names the structural shift. Dorsey shows what it looks like in practice. The experiment proves the mechanism. They are describing the same thing at four different levels of resolution.
Why 94% of AI Transformations Fail —
and Why It's Not About the Tech
My core thesis on organizational transformation runs through three layers that must move simultaneously: Technology, Organization, and Human. The Harvard/Stanford experiment ran a controlled test and confirmed the model.
The 94% GenAI pilot failure rate is not a mystery. It is the predictable output of organizations treating AI transformation as a technology deployment while leaving the organizational and human layers structurally unchanged.
Post-mortems on failed AI pilots return six root causes, repeatedly, across industries and geographies: lack of skills, high costs, inadequate tools, complex projects, data complexity, confidence gap. These are not six problems. They are one problem with six faces.
is knowledge distance, named from a different vantage point.
There is a compounding problem the experiment doesn't fully name. I've called it the Agentic Jevons Trap: when AI access expands without closing the knowledge distance, organizations generate more AI activity without generating more value. Jevons observed in the 19th century that improved steam engine efficiency increased coal consumption — not decreased it — because it made steam economically viable for more applications. The same dynamic fires in AI deployment when the human layer isn't ready. More agents. More throughput. More degraded output shipped at scale. The trap doesn't look like failure. It looks like productivity — until the output lands.
What Specific Knowledge
Is Actually Worth
Naval's specific knowledge framework reframes something that most organizations have systematically undervalued: the practitioner who accumulated domain knowledge over decades of real exposure, who knows not just the rules but the exceptions, not just the system but the reason the system works the way it does.
The experiment proves why this person is irreplaceable at the execution layer. AI can equalize conceptualization — structuring problems, generating options, building outlines. That gap is effectively closed and will continue to close. But the execution layer — judging AI output against deep domain knowledge, catching errors that only context can reveal, making the call a model can't — requires exactly what the domain practitioner carries.
As AI equalizes conceptualization, the strategic value of domain expertise concentrates at the execution layer — exactly where knowledge distance determines whether output gets elevated or degraded. The practitioner who can evaluate AI output is not competing with AI. They are the human layer that makes AI deployable in production.
The market has invented a job title for this person.
Forward Deployed Engineer: someone who deeply understands a domain, earns the trust of senior practitioners, and makes AI work in real environments — not demos. The market invented that job because organizations hit the wall the experiment just measured, at scale, and had to find a name for the person who closes it.
The governing principle is revenue per employee, not headcount reduction. The organizations that understand this aren't replacing domain practitioners with AI. They're encoding what those practitioners know into governed, auditable systems — and moving those practitioners up the stack to the judgment layer where their specific knowledge compounds rather than depreciates.
Two Futures Are Being Built
Right Now, Simultaneously.
Both futures are under construction inside organizations that don't yet know which one they're contributing to. The decision point isn't a strategic planning meeting. It's the next deployment decision.
Org never redesigned.
- —Agents deployed before knowledge distance is closed — autonomous execution on a foundation that can't evaluate output
- —Degraded decisions at scale, running unmonitored in production systems
- —The Jevons Trap fires — more AI activity, not more AI value
- —Stanford: 362 AI incidents in 2025, up 55% — this is what that count is made of
- —Practitioners sidelined; domain knowledge walks out unencoded
- —Hierarchy intact — now managing AI coordination overhead instead of human coordination overhead
moved in tandem.
- →Knowledge distance closed before agentic deployment scales
- →Practitioners at the governance layer — strategy, judgment, escalation
- →$172B in consumer value — Stanford's measure of what AI captures when it lands on a ready org
- →Hierarchy replaced by intelligence — Ismail's transition, completed
- →Domain knowledge encoded as auditable, governable system behavior
- →Specific knowledge compounds — the moat deepens with every cycle
The difference between these futures isn't the AI tools. Both organizations have the same models available. The difference is whether the organizational and human layers moved to meet the technology — or whether the technology was deployed into an org that wasn't designed to receive it.
The Wall Moves.
The Window Doesn't Wait.
The Harvard/Stanford researchers noted the GenAI wall is not permanent. As models improve, the knowledge distance threshold AI can bridge will shift. Capabilities that require domain proximity today may not require it in eighteen months. This is not a reason to wait. It is the strongest possible argument for moving now.
The organizations that respond to this finding by encoding practitioner knowledge into governed, executable systems — while the knowledge is still living in the people who hold it — are building a compounding asset. Every encoded cycle becomes harder to replicate. The moat isn't static. It compounds.
The organizations that wait are running a different calculation — one they haven't fully modeled. Not just attrition risk, though that risk is real and accelerating. They're in a window where the distance between the knowledge holder and the encoding process is still manageable. That window is closing. The gap between the two curves in that diagram is not hypothetical. It is the operating environment right now.
Naval, Ismail, Dorsey, and Harvard/Stanford arrived here from different directions. The conclusion is the same: specific knowledge is the asset, the org must be redesigned to capture it, and the time to do that is before the practitioner walks out the door — not after.