Two HR specialists are advising two different client companies. Both situations look identical: a mid-level employee has been underperforming for six months, their manager wants to terminate, and the company has a progressive discipline policy. Same surface features. Same policy. Same apparent answer.
A senior specialist looks at Situation A and says: proceed with termination, the documentation supports it. She looks at Situation B and says: slow down, this one is different. In Situation B, the employee was recently transferred from a department that was quietly toxic. The underperformance started after the transfer. The progressive discipline policy technically applies, but the deep structure of the situation — a systemic problem masked as individual performance — requires a different approach entirely.
An AI system, trained on policy documents and past cases, would match both situations to the same output. The surface features are identical. The AI has no way to see what the senior specialist sees, because what she sees isn't in the data. It's an abstraction she carries from years of cross-domain experience — a pattern she recognized between cases, not within them.
Surface structure vs. deep structure
Every problem has two layers. Surface structure is what's observable: the features, the data, the things you can point to. Deep structure is the underlying set of relationships, causes, and dynamics that generate the surface features. Two problems can have identical surface structure and completely different deep structure — or different surface structure and identical deep structure.
AI excels at surface-structure pattern matching. Given enough examples, it learns to map surface features to outputs with remarkable accuracy. This is genuinely powerful. Most of what knowledge workers do hour-to-hour is surface-structure work: looking up answers, matching situations to precedents, applying rules to cases.
But expert judgment operates on deep structure. The senior specialist didn't have a rule for "situations where underperformance masks systemic dysfunction." She had an abstraction — a deep-structural pattern she'd seen across different domains, different companies, different surface presentations. She recognized it in Situation B not because she'd seen an identical case, but because she'd seen the underlying shape before in a completely different context.
What this predicts
If you know how much of a role is surface-structure work versus deep-structure work, you can predict how AI will affect it.
Heavily surface-structure roles get dramatically assisted or automated. Customer support triage, first-pass legal review, standard medical coding — AI handles these well because the work is primarily about matching observable features to known outputs.
Heavily deep-structure roles are the ones where AI becomes a tool but can't replace the practitioner. Experienced consultants, senior diagnosticians, crisis negotiators — the value of these roles is in recognizing patterns that aren't in any dataset, because they're abstractions that exist in the expert's accumulated experience.
Most roles are a mix. This is what makes the "will AI replace X job?" question unanswerable in aggregate. The answer is always: it depends on the ratio. AI will handle the surface-structure portion and leave the deep-structure portion to humans. The job doesn't disappear. The composition of what the job is changes.
Why this matters
The policy conversation about AI and employment needs better tools for thinking about where AI's capability boundary falls. "AI can do anything" is wrong. "AI will replace everyone" is wrong. "AI won't affect my job" is also wrong.
The surface/deep distinction offers something more useful: a way to ask, for any specific role, which parts are surface-structure work that AI can pattern-match, and which parts require the kind of deep-structural reasoning that only comes from years of cross-domain experience.
The answer matters for training, for organizational design, and for how we think about expertise development. If we automate the surface-structure work that used to be the training ground for developing deep-structure judgment, we need to ask: where does the next generation of experts come from?