Last week, researchers from Harvard Business School, Stanford University, and the Stanford Digital Economy Lab published results from a field experiment that has been running inside a major professional services firm. The findings named something that has been operating quietly inside every AI deployment — including the ones that are failing.
The experiment is simple in concept: give some knowledge workers access to AI assistance for their work, give others none, and measure the output quality gap. The finding everyone expected was that AI users would outperform non-users. The finding that matters is which AI users outperformed — and why the others didn't.
The variable that determined AI output quality wasn't the model. It wasn't the prompt. It wasn't the compute. It was proximity. The workers who produced the best AI-assisted outputs were the ones who already held deep domain expertise. They knew what good looked like, could evaluate what the model produced, and could redirect it when it drifted. The workers without that expertise — even with access to the same tools — couldn't close the gap. In some cases they widened it.
Domain expertise isn't just valuable in the AI era. It is the specific variable that determines whether AI output gets elevated or degraded. Harvard and Stanford just proved it.
This is the Knowledge Distance Problem.
The binding variable
KD is the root cause
Structural KD advantage
What Knowledge Distance Actually Is
Knowledge Distance is the gap between what AI requires to execute reliably and what the people inside an organization can actually provide, evaluate, and govern at the moment of deployment. It is not a skills deficit in the traditional sense. It is a proximity deficit.
The people closest to the work don't know how to translate their domain expertise into the machine-readable context that agentic systems need to make good decisions. The people responsible for governance don't understand what they're overseeing. The people who bought the tools assumed the tools would figure it out.
KD is why the AI adoption data keeps telling the same story regardless of the research organization doing the asking. 94% adoption, 6% bottom-line impact. 79% experimenting, 8.6% in production. 86% of CEOs say employees have the skills — 25% are actually using AI regularly. Every one of those gaps is a KD gap. The technology arrived. The distance between the technology and the people who could use it effectively didn't close automatically.
"The bottleneck was never intelligence — it was the translation layer between knowing and building. That layer is collapsing."
Andrej Karpathy · Former Director of AI, Tesla · OpenAIThe Three-State Misread
Most organizations are trying to execute a transformation they've misdiagnosed. They think they're moving from a Technology Upgrade Cycle to a Human Augmented state — better tools, smarter workflows, productivity gains layered on top of existing structures.
That's not the destination. That's the wall.
Most organizations get stuck in the Human Augmented state because they treat AI deployment as a tooling project rather than a transformation project. They add capability without redesigning authority, decision rights, or workflow logic. The tools arrive. The organization doesn't change. And so the tools underperform — not because they're weak, but because they're operating inside a structure that was never redesigned to use them.
The Six Root Causes That All Collapse to KD
Every standard explanation for why AI deployments fail is a symptom. The root cause is always Knowledge Distance.
The Agentic Amplifier
KD was always present in AI deployments. Agentic AI makes it dangerous at scale.
When AI generates a document or a summary, a human reviews it before it does anything. The KD gap shows up as review time, rework, or a poor output that gets corrected. Manageable.
When AI agents execute decisions autonomously — pricing, routing, approvals, workflow orchestration — the KD gap shows up as compounding errors at scale, unmonitored. The system is operating on domain knowledge it doesn't have. The humans who should be catching the drift don't have the proximity to see it happening.
KD × agentic deployment = degraded decisions at autonomous scale, unmonitored. This is the governance gap the IBM IBV identified when they found that organizations with AI deployed have AI-related security incidents rising 55% year-over-year. The incidents aren't a technology failure. They're a KD failure operating at agentic speed.
What This Means for IBM i Practitioners
The Harvard/Stanford finding has a specific application here that IBM i practitioners are underestimating.
IBM i organizations hold something most enterprises are desperately trying to acquire: 30 to 40 years of encoded domain expertise in production systems. The business logic embedded in RPG programs, the data governance baked into Db2, the workflow intelligence in the job stream and subsystem architecture — all of it is proximity, encoded. That is the specific variable Harvard and Stanford identified as the determinant of AI output quality.
Dario Amodei named it plainly: there's a large gap between AI that works in a demo and AI that works in a regulated industry. To close that gap, you need domain expertise. IBM i practitioners close that gap by definition. They don't need to acquire the domain knowledge. They are the domain knowledge.
What changes with KD as the named framework is the strategic implication. This isn't about IBM i being "good enough to modernize." It's about IBM i practitioners being structurally positioned on the right side of the binding constraint in AI deployment — before anyone else in their industry recognized it as a constraint.
Data proximity: 40 years of transactional data with referential integrity, domain-encoded in Db2. The model doesn't need to be trained on what the data means — the practitioner already knows.
Process proximity: The business logic in RPG and CL programs is the encoded institutional knowledge the agentic layer needs to make good decisions. It exists. It just needs to be surfaced and connected.
Platform proximity: IBM i's event-driven architecture, job stream model, and object-level governance are native analogs to the agentic execution patterns enterprises are building from scratch. The distance to agentic is shorter — because the platform was already there.
The Organizational Implication
KD is not primarily a technology problem. It's an organizational design problem. The IBM CEO Study confirmed it: the gap between capability and deployment is an organizational design problem, not a skills problem. 86% of CEOs say their employees have the skills to collaborate with AI. 25% are actually doing it.
Closing KD requires three things running simultaneously — what Signal4i tracks as the HOT framework:
Human: Domain experts need to be in the AI workflow loop as evaluators and governors, not replaced by it. Their proximity is the quality signal the agentic layer can't replicate.
Organization: Decision rights, handoff protocols, and governance frameworks need to be designed before agents are deployed — not retrofitted after they've already drifted. The 4× business objective delivery finding from IBM IBV is what organizational KD reduction looks like at scale.
Technology: The stack needs to connect the encoded domain knowledge (data, business logic, process architecture) to the agentic layer above it. For IBM i, that connection layer is now available: MCP Interface Layer, IBM Bob, Mapepire, the AI SDK for Db2.
The KD problem is solvable. IBM i organizations have the raw material — the domain proximity — that most enterprises are still trying to build. The move is to connect it deliberately to the agentic layer, with the organizational design to govern what runs.
"The constraint was never the data. The constraint was the distance between the domain and the model doing the work. IBM i practitioners are already there."
Signal4i · Field Note 04 · May 1, 2026