Summarise this page with your favorite AI assistant
I once gathered nearly a thousand employees from all of our divisions for a week of intensive training. We taught them to work within a single system: entering data into the CRM, logging attendance and sales where the entire company could see them, and stopping the habit of duplicating processes in personal Excel files. For a week straight, we repeated the same message: the system—yes; Excel—no. The feedback on the training was overwhelmingly positive. I was sure everything was about to change.
Two weeks later, I stopped by one of our branches and saw the manager entering data into her own Excel spreadsheet again. I asked why she had gone back to the old way. "It's just easier for me," she said.
My mistake was assuming that once people understood something, they would immediately start applying it. Today, many companies are falling into a similar trap with AI. They buy tools, hand out access, run training sessions, and assume that's enough. Swagatam Basu, Senior Director Analyst in the Gartner HR practice, called this the "enablement illusion"—when access and adoption metrics are mistaken for transformation.
I'm convinced the problem isn't that people are bad at prompting or resistant to change. So what is it?
Companies have poured enormous sums into AI, and most have yet to see a return. Gartner reports that 88% of HR leaders saw no meaningful business value from AI tools in 2025 [1]. A BCG study of 1250 companies found that only 5% are capturing significant value from AI at enterprise scale, while 60% are seeing no material return at all [2].
The default reaction is to blame the technology: wrong model, weak governance, disconnected tools. And all of that can genuinely get in the way. When AI systems are siloed and data doesn't connect, the model lacks context. Orchestration—a separate layer that ties fragmented tools together and consolidates their data into a unified picture—can help here. But that solves the problem at the technical level. Even perfectly integrated systems will only pass along the knowledge a company has managed to capture.
As someone who has spent over twenty years in education and employee development, I've seen this firsthand: the way your best people actually make decisions is almost never documented anywhere. The consequences are slow onboarding, knowledge lost with every departure—and now, AI investments that don't pay off.
AI only scales knowledge that a company has made explicit. If the model receives nothing but formal documentation describing an ideal process rather than the real one, it will produce answers that merely look correct. Employees then either spend hours manually verifying the AI's recommendations or simply stop using the tool altogether.
The kind of knowledge I'm talking about has a precise name: tacit knowledge—what a person knows how to do but cannot break down into steps. And the more experienced the expert, the harder it is for them to do so. Over years of practice, their knowledge has become automatic: they make the right call without consciously knowing how they got there. That's why "just ask the expert how they do it" doesn't work—they genuinely don't know the answer.
This is an old and expensive problem. APQC (American Productivity & Quality Center) data shows that 85% of senior executives are concerned about knowledge loss when experienced employees leave, yet only 8% of organizations consistently preserve that knowledge.
The time lost searching for knowledge didn't start with AI either. According to other APQC findings, professionals spend roughly 8 hours a week hunting for the information they need or re-explaining things they've already explained before. That's nearly an entire workday per week that a company loses simply because knowledge isn't available in a transferable form.
Onboarding, which depends directly on effective knowledge transfer, tells the same story. Gallup reports that only 12% of employees strongly agree their company does a good job onboarding new hires, and reaching peak productivity typically takes a new person about a year. If a company can't quickly and fully transfer a role to another human being, it's unlikely to transfer it to AI.
I've noticed that organizations that learned long ago to capture and structure the knowledge of their best people find AI adoption significantly easier today—they actually have something to feed it. Those that never learned blame the model.
Many companies already have everything they need for successful AI adoption—except knowledge of how their best people actually work. I look at this through the lens of my background in educational science: extracting knowledge and turning it into a transferable skill is a classic learning problem. Here's how to approach it:
After the Excel incident, I stopped running training sessions. Instead, I sat down with the best performers and began, step by step, to break down how they actually make decisions: what criteria they use, what they discard, where they deviate from the playbook. Over time, this became the foundation of our hiring and training system—and the extracted knowledge became the base on which our AI operates. It answers questions and trains people drawing on the real experience of our best, including my own, rather than on generic instructions.
When knowledge stopped being the personal property of individual people, everything changed. I was able to step out of day-to-day operations, and new divisions started opening without my involvement—in a couple of weeks instead of months. Not because we found a magic tool, but because we stopped losing the knowledge of our best people.
If your last AI tool didn't meet expectations, don't rush to test the next one. Ask yourself a few questions instead:
There's a good chance you won't like the answers. But if you don't ask yourself these questions, the invoice for the AI tool that never paid for itself will ask them for you.