I worked on a project that deployed an AI solution to 309,000 worksite employees at a large professional employer organization. I expected it to democratize expertise. The tool was simple: semantic search over HR policy documents. Instead of calling an HR specialist or digging through a document portal, any employee across hundreds of client companies could ask a question in natural language and get the right policy.
It worked. Search abandonment dropped significantly. People who couldn't find information before now could. Straightforward.
We also built a retrieval system for HR specialists who advise those client companies. Same underlying technology. Completely different effect. The specialists didn't become less needed. Their routine recall was automated — they no longer had to remember which policy lived where. But the actual hard part of their work was untouched: interpreting edge cases, reading organizational context, advising clients through ambiguity.
Same tech. Two populations. Opposite dynamics. I've been thinking about why ever since, and I think transfer theory offers the clearest explanation.
Two kinds of knowledge transfer
Cognitive scientists distinguish between two modes of knowledge transfer. Low Road transfer is automatic and routine — pattern matching you do without conscious effort. A specialist who's looked up the parental leave policy a hundred times doesn't think about it. They just know where it is. High Road transfer is conscious and effortful — the kind of reasoning that requires you to recognize deep structural similarities between situations that look different on the surface. A specialist advising a client through an unusual termination scenario is doing High Road transfer: drawing on years of cross-domain experience to navigate something the policy manual doesn't quite cover.
AI is extremely good at Low Road transfer. Given enough examples, it can match queries to answers, surface relevant documents, and automate routine retrieval. That's what the employees experienced — AI handled the Low Road work they used to struggle with.
But AI doesn't transfer nearly as well on High Road tasks. The specialist who recognizes that two situations — identical on paper — require opposite advice because the underlying organizational dynamics are different? That's High Road reasoning. It requires abstraction, analogy, and years of accumulated judgment. No retrieval system can automate it because the knowledge isn't in the documents. It's in the specialist's head.
The expertise gap widens
Here's the counterintuitive implication: AI didn't close the gap between junior and senior specialists. It made the gap more visible.
Before AI, part of what separated a senior specialist from a junior one was Low Road knowledge — the senior person just knew where everything was. AI erased that advantage. Now anyone can retrieve information instantly. But the gap that remains is entirely High Road — judgment, pattern recognition across cases, the ability to see when a situation that looks routine is actually novel. That gap is harder to close because it can't be learned from documents. It comes from years of engaging with ambiguity.
For the employees who gained access to information they couldn't find before, this is purely good. AI solved a friction problem. For the specialists, it's more complex. Their Low Road work was devalued. Their High Road work became more important. The composition of what makes them experts shifted — and that shift isn't captured by asking "did AI replace jobs?"
What this means
The employment impact conversation tends to ask aggregate questions: how many jobs will AI eliminate? What will happen to wages? Those questions treat "AI deployment" as one thing. What I observed was that a single deployment produces different dynamics depending on where you sit in the organization — reducing friction for one population while restructuring expertise for another.
Transfer theory predicts this. If you know which kind of knowledge transfer a role relies on, you can predict how AI will affect it. Low Road roles get automated or dramatically assisted. High Road roles get redefined — their routine substrate disappears, leaving the conscious, effortful judgment exposed as the core of what they do.