Prukalpa Sankar of Atlan wrote the line that should end every agentic AI post-mortem: "A human hire gets six months of structured exposure to the business before they're trusted to act on their own. An AI agent gets a prompt." She's right. But she frames this as an AI integration problem. I think it is a documentation debt problem that AI made impossible to ignore.
The test nobody ran
The agent extension test asks whether you can describe how you think clearly enough that an agent would apply it to a new case and you'd endorse the result. If yes, your thinking is a method. If no, it's a habit. Most organizations have never run this test on their own operations. The agent's context failure is the result arriving uninvited.
Sankar identifies three gaps in how agents encounter organizations: connectivity, semantics, and institutional knowledge. The first two respond to better infrastructure, better data pipelines and schema documentation. The third doesn't respond to any model upgrade.
Bakal's research on agentic software development states the bottleneck plainly: "The bottleneck to effective agentic software development is not model capability but knowledge architecture." When a team says the agent doesn't understand their business, they're reporting a knowledge architecture failure. The agent follows written policy and misses unwritten exceptions, because those exceptions were never documented.
The debt predates the agent
Ninety-two percent of organizations fail to consistently capture knowledge from retiring employees. Deloitte estimates that retirement knowledge loss costs $6.9 to $9.6 trillion in economic output. Those numbers existed before any agent was deployed.
The knowledge was already gone. Humans compensated with hallway conversations and tribal memory. Agents cannot.
Only 39% of organizations attribute any EBIT impact to AI. Gartner projects that over 40% of agentic AI projects will be canceled by end of 2027. S&P Global found that companies abandoning most AI initiatives doubled from 17% to 42%. McKinsey reports only 1 in 10 agent pilots move beyond pilot stage. Documentation debt is the common cause wearing a technology label.
Bakal describes the downstream mechanism: each cycle of agent guessing, failing, and retrying fills the context window with the detritus of failed interactions and degrades the agent's reasoning. The context that would prevent the cascade was never written down.
Where Sankar is right, and where I push further
I agree that onboarding is the right frame. Sankar observes that the model is functionally blindfolded. Atlan's benchmark found a 38% improvement in text-to-SQL accuracy when agents operated with governed institutional context. That gain came from documentation that finally existed.
The California Management Review argued that the real differentiator for enterprise AI is the tacit knowledge embedded in people's judgment. That framing explains why model upgrades keep disappointing: they improve reasoning on visible data while the invisible data stays invisible.
I push further on the implication. Sankar treats the onboarding gap as something AI teams should close. I think the gap belongs to the entire organization. It was never closed for human employees either.
The senior engineer who just knows how the billing exception works absorbed that knowledge over years of proximity. When they leave, the knowledge leaves. Agents just made the loss instant and visible.
Meta's engineering team described what happens without that context: agents that guess, explore, guess again, and often produce code that compiles but is subtly wrong. That describes the new hire who follows the handbook and misses every unwritten rule.
IDC reports that 89% of organizations have ongoing data-quality problems. The next model will not fix this. Your agent's context failure is a receipt for institutional knowledge the organization never made explicit. Count how many critical processes in your organization exist only in someone's head. That is your AI readiness score, and no model upgrade will change it.
.webp)
.webp)
.webp)
