Claude Took 80% of a Researcher’s Questions, Not His Job

Claude absorbed most of a recurring support duty without replacing the researcher who performed it. In his account published December 8, 2025, early Anthropic research hire Andy Jones wrote that 80% of the new-hire questions reaching him had disappeared within six months, while emphasizing that answering them was only one part of his job.

Evidence published in 2026 makes the underlying warning more relevant but also more precise. AI can cross the threshold for automating a defined workflow abruptly, yet the available research still shows uneven effects across tasks and occupations rather than a single timetable for replacing workers.

What actually changed in Jones’s work

Jones and other experienced employees had been answering technical questions from new hires. Once Claude became useful enough for that support function, employees could redirect routine questions to the system instead of relying on scarce researcher time.

The distinction is important: Claude surpassed the human group on a bounded activity, not across every responsibility attached to a research position. The case demonstrates task substitution—a recurring unit of work moving from people to software—rather than the documented elimination of a complete job.

This kind of change need not track model improvements smoothly. A system below an organization’s acceptable threshold may attract little use; a modest gain in reliability, accessibility or cost can then make it the default channel. Adoption appears sudden even when the technical progress behind it was cumulative.

Why the horse analogy is useful

The metaphor captures the gap between gradual technical improvement and abrupt practical consequences. Employers do not adopt “intelligence” as an abstract quantity; they adopt systems that can resolve a question, revise code, produce a document or complete another identifiable task under acceptable conditions.

A capability improvement becomes economically significant when it reduces the supervision, retries or specialist intervention required to obtain a usable result. At that point, an established workflow can switch quickly even if the surrounding occupation remains intact.

The analogy breaks down when a worker is treated as equivalent to a single output. Most occupations combine production with prioritization, accountability, negotiation, institutional knowledge and work in physical or social environments. Removing one recurring duty can leave the role largely intact, reshape it around judgment or, in some cases, narrow it enough to affect future hiring.

Later evidence points to wider delegation

The June 26, 2026 Anthropic Economic Index report found that Claude sessions increasingly involved long-running agentic tasks and that more than one-third of surveyed users expected AI to perform most or nearly all of their work tasks within the following 12 months; it also warned that the sample was not representative of the workforce and disproportionately included computer, mathematical and management occupations.

That limitation prevents the survey from serving as a general forecast. Its value is narrower: expectations of expanding task coverage appeared across a broader population of active Claude users, while respondents who delegated more work also tended to express greater optimism about their employment outcomes. Greater automation and greater confidence can therefore coexist.

The relationship does not establish that delegation causes optimism. Enthusiastic users may simply be more willing to hand over complete tasks, and self-assessments cannot show whether skills or employment prospects will improve in practice. The findings describe current users’ behavior and expectations, not a measured rate of occupational replacement.

Exposure is not the same as displacement

The IMF study published July 13, 2026 used occupational microdata to conclude that most Israeli workers were likely to benefit from AI adoption, while about one-fifth faced high exposure combined with low complementarity and therefore greater displacement risk.

The estimate is specific to Israel and cannot be treated as a global rate. It nevertheless illustrates why exposure must be separated from complementarity: two occupations may both contain many AI-compatible tasks, yet one may gain productivity from the technology while the other loses a larger share of the work that justifies its staffing.

Job titles alone obscure this difference. A researcher, administrator or software developer may perform some activities that are easy to specify and verify alongside others that depend on tacit context, responsibility or relationships. The balance among those activities matters more than whether the occupation is broadly classified as exposed.

The workflow is the relevant unit of change

A rapid transition is more plausible when work arrives frequently in a repeatable form, the necessary information can be supplied to the system and a user can recognize an acceptable result without recreating the entire process. Reversible errors and a clear route for escalating unusual cases also lower the cost of delegation.

Work with ambiguous objectives, fragmented confidential context or costly mistakes creates more friction. AI may still assist with drafts or analysis, but human involvement remains harder to remove when someone must interpret competing goals, negotiate with others or accept responsibility for the outcome.

This is why employment totals may register change later than workflows do. An organization can automate internal support, routine revisions or first drafts without immediately eliminating a position. Earlier effects may include fewer entry-level assignments, altered hiring criteria, higher output expectations and a greater concentration of work in review and judgment.

The gallop has no universal starting line

The central warning survives scrutiny: practical displacement can feel abrupt after years of apparently incremental progress. What the evidence does not support is the stronger claim that every profession is approaching the same point or that automating a large task necessarily eliminates the worker who previously handled it.

The “gallop” begins separately for each workflow, depending on capability, available context, cost, accountability and tolerance for error. Jones’s experience shows how quickly one channel of work can change; the later research shows why the consequences still range from substitution to complementarity rather than resolving into a single future for all workers.

Claude Took 80% of One Researcher’s Questions—Not His Whole Job