AI Will Make Expertise Cheap—and Expose How Much of It Was Gatekeeping

AI will make a great deal of professional expertise dramatically cheaper, and people in prestigious knowledge-work fields should stop treating that as a distant possibility. It will not make lawyers, doctors, engineers, investors, or consultants disappear—but it will force those professions to confront how much of what clients bought was expensive production of analysis rather than uniquely human judgment.

This is not the usual claim that AI will “assist” professionals. Assistance is already the conservative framing, useful because it lets firms preserve the old org chart while adding a chat box. My claim is more disruptive: once a competent analytical first draft, research plan, scenario analysis, implementation sketch, and critique can be generated cheaply and iterated quickly, the economic floor under a huge amount of expert work falls. A junior associate’s memo, a consultant’s market map, an engineer’s migration plan, a financial analyst’s sensitivity model, or a clinician’s differential diagnosis will still require review. But the first-pass intelligence that once took hours or days will increasingly cost something closer to compute plus context.

The evidence points to capability expansion, not just faster typing

The early workplace evidence matters precisely because it is less dramatic than the marketing. A widely cited field study of more than 5,000 customer-support agents found that access to generative-AI guidance increased issues resolved per hour by roughly 14%, with larger gains for less experienced workers. That is not a robot replacing a department. It is more consequential: the tool transfers patterns embedded in the work of stronger workers to people who have not yet accumulated those patterns themselves.

Experiments with consultants show the same mechanism in a more recognizably white-collar setting. On selected tasks within the model’s competence, AI users completed more work, worked faster, and produced higher-quality outputs; on tasks outside that competence, the system could confidently steer them wrong. The important lesson is not that consultants have been solved. It is that a general-purpose model can give a capable non-expert a temporary, task-specific skill boost. That is the beginning of cheaper expertise.

Software engineers should recognize the pattern immediately. We already know that a good codebase search, a test suite, a linter, a debugger, and a senior review turn broad programming knowledge into deployable change. An LLM adds a fast, unusually broad collaborator to that pipeline. It can translate an issue into candidate implementation paths, locate relevant modules, generate test cases, explain an unfamiliar subsystem, and draft the tedious parts. That does not eliminate the need to understand invariants, production constraints, operational failure modes, or the business reason for the change. It does make the path from “I do not know how to start” to “I have something reviewable” much shorter.

Every expert field has a cheap layer and an accountable layer

Law is the clearest example. Much legal work involves finding authority, comparing clauses, extracting facts from a record, drafting a first version, and explaining options in ordinary language. Those tasks are linguistic and document-heavy—the exact terrain where modern models are useful. The result should be cheaper access to routine legal help, especially for people and small businesses that currently receive no help at all. But filing a brief with invented cases is not a harmless formatting bug. The lawyer remains accountable to the court, the client, professional rules, and the actual jurisdiction. AI cheapens the work product; it does not cheapen responsibility.

Medicine has an even sharper boundary. Models can synthesize a chart, surface relevant evidence, suggest questions, produce a differential, and reduce documentation burden. Recent clinical-reasoning evaluations show genuine progress, but a benchmark vignette is not a patient with incomplete history, conflicting values, fragile trust, messy follow-up, and a body that may not resemble the training distribution. The physician’s role becomes less about being a walking recall database and more about integrating evidence with examination, uncertainty, consent, triage, and accountability. That is not a demotion. It is a demand that medicine price and organize itself around the work only a clinician and care team can truly do.

Finance and consulting will be hit hardest where they sell analysis as a scarce artifact. A model can build a competitor brief, summarize earnings calls, generate a downside case, identify assumptions worth testing, or turn a pile of spreadsheets into a client-ready narrative. It cannot make a portfolio manager bear the loss, a CFO sign the forecast, or an executive make a politically costly reorganization succeed. Yet too much of the billable pyramid exists to manufacture slides, research, and plausible synthesis on the way to those decisions. Clients will notice when the production cost collapses.

The strongest counterargument: expertise is embodied, social, and accountable

The best objection is that this argument mistakes fluent output for expertise. Real expertise is not a memo or diagnosis-shaped answer. It is tacit knowledge built through feedback, judgment about what information is missing, skill at noticing when a situation is abnormal, credibility with stakeholders, ethical duty, and liability when things go wrong. A lawyer persuades and negotiates. A doctor examines and earns trust. An engineer navigates organizational dependencies and gets paged at 3 a.m. A financial adviser understands a client’s constraints. A consultant gets a divided leadership team to act. In this view, AI may commoditize junior busywork but leaves the valuable core intact—and may even make senior experts more valuable.

That objection is correct about the core and wrong about the conclusion. Expertise has always included judgment and accountability, but many professions bundled those scarce qualities with a large volume of analysis that was merely labor-intensive. The bundle was defensible because producing competent analysis required trained people and time. AI breaks that coupling. A senior professional may indeed become more productive and more valuable in absolute terms. But the number of expensive human hours required before a client reaches that professional should fall. The professional who survives unchanged will be the one whose price was mostly a toll for accessing analysis.

There is also a training problem we should not dismiss. Junior work is how people learn. If AI consumes the repetitive research, drafting, modeling, and coding tasks, firms could hollow out their apprenticeship pipeline and discover too late that they have no seasoned reviewers. But that is an argument for redesigning training, not for preserving artificial scarcity. Give juniors AI-assisted work with explicit verification duties, compare their reasoning against source material, rotate them through real decisions, and make them explain why a recommendation fails—not merely how to produce a polished one.

What should change now

Professional firms, hospitals, engineering organizations, and regulators should stop measuring AI adoption by seats purchased or documents generated. They should unbundle work into: machine-generated analysis, human verification, accountable decision-making, and relationship work. Then they should price, train, audit, and staff each layer honestly.

For software teams, that means treating AI output as an inexpensive proposal rather than earned truth: require tests, provenance for claims about unfamiliar systems, reviewable diffs, and clear ownership of production outcomes. For law, medicine, finance, and consulting, it means building workflows that make expert review visible rather than silently using a professional’s name as a warranty stamp on machine output. And for everyone who sells expertise, it means lowering prices where the work has become cheap. If AI makes competent analysis abundant, the public should receive more than faster deliverables; it should receive access to expertise that was previously priced out of reach.