# AI Makes Expertise More Valuable Before It Makes It Less Valuable Source: https://www.bobzhu.tech/ai-makes-expertise-more-valuable/ Markdown: https://www.bobzhu.tech/assets/agents/ai-makes-expertise-more-valuable.md Tags: Essays, AI, Work, Systems Summary: Current agents lower implementation barriers while rewarding the domain knowledge needed to specify work, catch edge cases, and recover from failure. Feature image: https://storage.ghost.io/c/ea/80/ea80b01b-c9d0-45fd-a95f-5fbcd52ed925/content/images/2026/08/expertise-leverage-feature.png Feature image alt: Hand-drawn editorial illustration of expert constraint cards steering an execution machine through a maze. Feature image caption: The agent can run the route. Expertise still determines which route is worth running and where it can fail. The common version of the AI-and-expertise story runs like this: AI will eventually automate everything, so domain expertise loses value. Hold on long enough and the knowledge you spent years building becomes a commodity. That story might be right eventually. But it gets the near-term direction backwards. Right now, AI agents reward domain expertise more directly than most of the work environments that came before them. The reason is structural. AI removes implementation barriers, but it shifts the cognitive weight of work toward specification, judgement, edge-case recognition, and recovery from failure. Those are the parts of work that expertise has always owned. ## What AI has actually removed Implementation barriers were always friction, not the core of a domain. Writing code to query a database, generating a first-draft synthesis of twenty papers, converting a rough process map into a structured document: these tasks required technical skill, but they were not the source of judgement. They were the path toward it. AI handles much of that path now. A skilled analyst who previously spent two hours writing SQL can direct an agent to do it in minutes. A researcher who needed a week to process a literature review can compress that substantially. Work that required translation from thought to output has shortened. This matters because it accelerates access to the part where the hard thinking lives. You arrive at the interesting problem faster. Arriving faster does not reduce the value of being able to recognise the problem when you get there, or knowing what to do about it. The bottleneck moves. It does not disappear. ## The cognitive work that stays in the human lane AI does not remove knowing what to ask for, noticing when the output is wrong, defining what "done" looks like, and recovering when things go sideways. These are not soft skills or vague professional virtues. They are specific cognitive operations, and they correlate directly with domain expertise. To specify a task well, you need a mental model of the problem space, an understanding of which edge cases matter, which constraints are real versus assumed, and which shortcuts introduce risk downstream. A novice prompt produces a plausible output. An expert prompt produces the right output. The expert also spots the difference between the two, which the novice may not. To catch errors, you need a reference for what correct looks like. AI systems make confident mistakes. They return syntactically clean but semantically broken outputs: code that runs but does the wrong thing, summaries that read well but misrepresent the source, plans that sound sensible but skip a critical constraint. Without domain knowledge, there is no filter between "the model says this looks right" and "this is actually right." Recovery is the clearest illustration. When an agent gets stuck, a tool call fails, a file arrives in an unexpected format, or the logic hits an edge case it was not prepared for, someone has to diagnose the problem and redirect. That diagnosis is domain knowledge applied to an unexpected situation. It cannot be delegated back to the agent that just failed. Each of these operations scales with expertise. The more you understand the domain, the better your specifications, the sharper your error-detection, the faster your recovery. The underlying knowledge base has not become less relevant. It has become the bottleneck. ## What a large-scale study suggests Earlier this year, [Anthropic published preliminary findings](https://www.anthropic.com/research/claude-code-expertise) from an analysis of roughly 400,000 Claude Code sessions involving around 235,000 people. The results deserve careful reading and appropriate caution: the researchers treat them as preliminary, the context is coding specifically, and the findings should not be generalised beyond what the study shows. With those limits noted, the patterns are interesting. In a typical session, people made roughly 70% of planning decisions, while Claude handled roughly 80% of execution decisions. The split is not incidental. It reflects where the cognitive weight sits in agent-assisted work: humans set the direction, agents run the steps. Domain expertise correlated with meaningful differences in outcomes. Experts extracted more work per instruction, recovered from problems more effectively, and achieved higher verified success rates. The largest gains appeared moving from novice to intermediate, not from intermediate to expert. Even moderate domain knowledge substantially changes what you can extract from an agent. The floor matters more than the ceiling. One finding is worth pausing on. Coding sessions from people in major non-software occupations landed close to software practitioners on success measures. Domain specialists from other fields who used Claude Code within their area of expertise performed more like professional software developers than non-expert coders. Expertise in a domain, applied to agent-assisted work in that domain, transferred. Preliminary findings from a specific context. But directionally they point the same way: domain knowledge shapes what you can extract from an agent, and that relationship holds across fields. ## Expertise changes shape, not quantity Before AI agents, expertise showed up in execution. The expert wrote the code, produced the analysis, drafted the document faster and better than the novice. Implementation was where the gap was visible. With AI agents, execution gets distributed. Claude writes the code. The gap now shows up in specification, verification, and recovery. The expert still outperforms the novice, but the performance difference lives in a different part of the process. This is expertise changing shape, not disappearing. The knowledge base that lets you specify well, validate correctly, and recover effectively is the same knowledge base you built through years of direct practice. It has not been automated. It has been moved to a more visible position in the work, where it now determines what gets done rather than how. There is a reasonable argument that future AI systems will handle specification and validation too. That argument is worth taking seriously. But current agents still require a human to close the loop, and the humans who close it well are the ones who understand the domain. For now, that remains the constraint. If you are building domain expertise, its relevance has not decreased. The mode of its application has changed, and for the moment, in the direction that rewards depth.
Working view

Where expertise moves

Execution gets cheaper, so the expert contribution becomes easier to see.

Layer What it carries
Specification Define the real task and its constraints.
Recognition Notice when plausible output is wrong.
Recovery Diagnose trouble and redirect the agent.
Completion Decide what evidence makes the work finished.
## Try this prompt Paste the following into Claude Code, Cursor, or a similar agent, replacing the bracketed sections with your own domain and context: "I need to [describe the task]. Before you start, list the assumptions you're making and the edge cases most likely to cause problems. Tell me what a successful output looks like and what would make you uncertain whether you'd hit it. Then proceed." The point is not the output from the first step. The point is reading the agent's stated assumptions against your own domain knowledge to catch the gaps before they compound. Expert specification starts here. ## Related on this site - [Your job isn't disappearing – parts of it are getting cheaper](https://www.bobzhu.tech/your-job-isnt-disappearing-parts-of-it-are-getting-cheaper/) - [The AI Function Clock](https://www.bobzhu.tech/the-ai-function-clock/) - [The AI Proficiency Ladder](https://www.bobzhu.tech/the-ai-proficiency-ladder/) ### Sources - Anthropic, "Agentic coding and persistent returns to expertise", 16 June 2026: https://www.anthropic.com/research/claude-code-expertise