AI Can Make Expertise Less Scarce

Skagway’s work begins from the concern that consequential expertise can become concentrated and difficult to transfer. That concern can become a habit of seeing scarcity everywhere. The more intellectually honest question is which parts of expertise are already becoming easier to distribute.

Generative AI supplies a serious answer in at least one domain. A large field study found that an AI assistant helped less-experienced and lower-skilled customer-support workers improve more than their experienced peers. If a tool can place useful patterns into the workflow at the moment of need, some benefit once tied to tenure can travel without a traditional apprenticeship.

That result does not weaken the case for studying expertise. It improves the diagnosis. An organization should not invest in preserving a scarce capability before asking whether the relevant part has become reproducible, searchable, or embedded in a tool.

Brynjolfsson, Li, and Raymond studied the staggered introduction of a generative-AI conversational assistant among 5,179 customer-support agents. The tool offered suggested responses and links to internal documentation during text conversations with customers. Productivity was measured as issues resolved per hour, a concrete operating outcome that combines handling time and resolution.

Access to the assistant increased productivity by 14 percent on average. The reported effect was heterogeneous: novice and lower-skilled workers improved by 34 percent, while experienced and highly skilled workers saw minimal impact. The authors also reported evidence concerning customer sentiment, employee retention, and possible learning, though the productivity pattern is the most relevant result here.

The design used a real workplace rollout rather than a short artificial task. It still concerned one company, one class of customer-support work, and one system trained in part on the organization’s historical interactions. Scope is central to the interpretation.

The authors provide suggestive evidence that the system disseminated practices associated with more able workers and helped newer workers move down the experience curve. They do not claim to have extracted a complete model of expert reasoning. The assistant could surface language and responses fitted to recurring conversational situations where quality and resolution were observable enough to learn from historical data.

This is a meaningful form of expertise distribution. A newer agent no longer needed to remember every relevant policy, search manually for every answer, or invent language from scratch. The workflow made an experienced pattern available close to the decision point.

The compression was not uniform. The minimal effect among the most experienced workers suggests the tool supplied less that was new to them—or that the measured outcome left less room for gain. It also warns against assuming an average productivity increase describes everyone’s work experience.

Customer support in this study had repeated interactions, textual records, internal documentation, and outcomes that could be measured. Those features make the work unusually amenable to a system that retrieves and generates from prior examples. The conversation still required human interaction, but much of the useful pattern was represented in data the tool could use.

Many executive decisions have thinner samples, shifting objectives, private information, contested outcomes, and consequences that emerge years later. Relationships may change what can be said; authority changes what can be done. A result from customer support should not be stretched across those differences.

The right inference is conditional: where work is repeated, well represented, and evaluable, AI may make parts of expertise less scarce. Organizations should identify those parts before treating the entire role as an irreducible reservoir of judgment.

The study supports claims about assisted productivity in a particular customer-support environment. It is compatible with easier distribution of response patterns, documentation access, language, and routine problem handling. It does not isolate each component or prove that workers acquired a durable unaided capability.

The study did not test CEO succession, rare strategic decisions, board governance, high-consequence exceptions, or the transfer of personal trust. It did not establish how performance would change after the tool was removed over a long period, or whether a workforce could evaluate failures outside the system’s learned patterns.

Those limits do not diminish the result. They stop it from becoming a slogan. A good operating decision asks which measured capability improved, for whom, under what conditions, and whether the organization still needs an unaided or independent version of it.

Before preserving a critical person’s work, separate retrieval, routine production, pattern matching, exception handling, relationship judgment, and decision authority. AI may already handle some layers well enough to reduce the dependency. Other layers may become more important because someone must define the problem, evaluate unusual output, or own the consequence.

This is an empirical question, not a permanent boundary between people and machines. Tools, tasks, and organizations will change. Annual review is more credible than a fixed claim that expertise will always remain scarce or inevitably disappear.

This article carries no service recommendation. Its purpose is to keep Skagway’s own premise falsifiable: preserve consequential human capability where evidence shows real dependency, and welcome evidence that makes a capability more broadly available.