Why Enterprise AI Fails
The Hidden Cost of Key-Person Dependency, and Why AI Made It Worse
Key-person dependency means a company's critical judgment lives in a few senior experts, so capacity is capped by their calendars and every departure removes capability the company never owned. Generic AI made it worse: it handles commodity work but fabricates on expert questions, so the hard decisions route to the same few people while their knowledge keeps leaking into unapproved tools. The fix is capturing expert judgment as a company-owned, checkable asset.
What does key-person dependency actually cost?
Ask an executive who signs off on the hard calls in their company and you get two or three names. Everything important routes through those names: the proposal that commits the firm, the design deviation, the portfolio anomaly. Most executives describe this as a hiring problem, “we need to grow the team.” It is more precise to call it what it is: the company’s core capability exists as a handful of employees’ private property, rented by salary, and the ceiling on growth is their combined calendar.
The costs compound quietly. Revenue is capped by review capacity, so the company turns down or delays work its juniors could execute if they could be trusted without the senior pass. Onboarding takes years because the only transfer mechanism is sitting next to the expert. Institutional memory fails: problems solved in 2019 are re-solved expensively in 2026 because the person who solved them left, and nobody can find or interpret what they left behind. And in the background sits the risk executives lose sleep over but rarely quantify: the resignation that would remove a capability, not a headcount.
None of this appears on a balance sheet, which is exactly the problem. The judgment of a company’s best people is usually its most valuable asset and its least owned one.
Why did AI make it worse?
AI was supposed to be the relief, and for generic work it was. But look at what the first wave of tools did to the expert bottleneck specifically.
The routine layer of work got faster: drafts, summaries, boilerplate. The expert layer did not, because on questions that depend on company-specific judgment, generic tools produce confident fabrications. The senior people caught those fabrications and rationally refused to let the tools near real decisions. Net effect: more junior throughput arriving at the same senior gate. The bottleneck did not move; the queue in front of it grew.
The research matches the experience. MIT NANDA found 95% of enterprise generative AI pilots delivered no measurable P&L impact, with integration into actual workflows and reasoning, the expert layer, as the failure point. In WRITER’s 2026 survey, 54% of C-suite executives said adopting AI is tearing their company apart, and only 29% saw significant organizational ROI. Deloitte’s 2026 work on pilot fatigue describes the executive response: disengagement by the third failed attempt.
And while the pilots failed, something worse happened informally. Employees, including the experts themselves, began pasting company material into whatever public tool helped them personally. Fragments of the company’s most valuable judgment now sit in chat histories the company does not control, training habits and expectations the company did not choose. The dependency remained; the moat started evaporating.
What does capturing expert judgment actually require?
Consider an anonymized composite we see constantly: a mid-size engineering firm where every proposal, before it goes out, is reviewed by one senior engineer with twenty years in the domain. He catches the under-scoped foundations, the client-specific constraint from a project years ago, the standard that changed in 2021. The firm wins work because of that review. It also cannot bid more than his calendar allows, and his retirement plans are a standing item in management meetings that no one puts in writing.
The naive capture attempts all fail the same way. Interviews and wikis produce documents about his judgment, not the judgment itself; six months later they are stale and unread. Handover periods transfer a fraction and take years. Fine-tuning a model on the firm’s documents bakes an approximation into one vendor’s product, uninspectable and uncorrectable.
What works is treating the judgment as an asset to be structured, not a story to be recorded. The firm’s source documents, its standards, past proposals, and project records, get organized into a knowledge library. Then, with the expert, the decisive material gets encoded deliberately: the contexts in which each standard applies, the catalog of failure patterns he checks against, the precedents behind his exceptions. This structured combination of sources plus judgment is what we call decision DNA, and building it is the heart of a knowledge and control layer.
“Every company has this person, and everyone in the room knows the name. The judgment in that one head is usually the most valuable asset the company has, and the plan for it is usually hope. Capturing how that person decides, in a form the company owns and can check, is the most valuable project most companies have never scoped.”
The Praxiron team
What changes when the judgment is captured?
The dependency inverts. Junior engineers get outputs grounded in the expert’s encoded judgment, each carrying a source reference and a calibrated confidence level, so the senior review becomes a check rather than a redo. Where the knowledge library does not cover a question, the platform abstains with “no sufficient source,” which is precisely the signal that a genuine expert decision is required; his calendar is spent only where nothing else suffices.
Onboarding compresses, because a new hire is working with the company’s judgment from week one instead of absorbing it by proximity over eighteen months. Institutional memory stops depending on tenure: the 2019 solution is in the library, referenced and findable, not in a departed employee’s recollection. And the retirement conversation changes character entirely. The expert’s knowledge becomes something he helped build into a company asset, still growing after he leaves, rather than something that walks out with him.
The capacity math is what executives notice first: the ceiling stops being the calendars of three people. That, not incremental productivity, is what “do far more without growing the team” means in practice.
If your version of this problem is currently called “we need to grow the team,” start by measuring it honestly, then read Activity metrics vs. outcome metrics for how to baseline the decisions that queue behind your experts, and Pilot purgatory for why another generic tool will not move that queue. When you are ready to see how judgment gets captured and put to work, see how Praxiron works.
Frequently asked questions
What is key-person dependency?
It is the condition where decisions that matter can only be made, or must always be reviewed, by a small number of senior experts. It shows up as one or two names on every hard escalation, proposals queuing for a single reviewer, and an executive who privately knows that a certain resignation letter would be a business event, not an HR event.
How do we identify our key-person risk?
Trace escalations for a month. List the decisions that stalled waiting for one specific person, the questions only one person could answer, and the reviews only one person could sign off. Then ask, for each name, what the company could no longer do the week after they left. The shortness of the name list and the length of the consequence list is the measurement.
Why did AI make key-person dependency worse instead of better?
Generic AI absorbed the routine work but fabricates on questions that depend on company-specific judgment, so exactly the expert decisions still route to the same few people. Meanwhile employees paste fragments of expert knowledge into public tools, where it trains nothing the company owns. The dependency stayed; the knowledge started leaking.
Can a company really capture an expert's judgment before they leave?
Documents alone cannot, which is why knowledge-transfer projects built on interviews and wikis disappoint. What works is encoding judgment as structured, machine-usable material: the contexts, standards, and precedents behind the expert's calls, connected to the company's source documents. Captured that way, the judgment works on every decision and stays when the expert goes.