If AI does the first draft, where do future experts learn?

Junior staff learn a great deal from work that experienced colleagues find repetitive. Researching a brief or preparing a first review gives them a view of the evidence before somebody more experienced explains what they missed. Automating the task can save time while removing the occasion for that conversation.

Anthropic’s How employment is changing at Anthropic, published on 16 October 2026, offers an unusually close view of that tension inside an AI developer. It should be read as a company-specific study, not a forecast of employment across the economy. The conditions at a rapidly growing frontier laboratory are far from typical.

For a marketing or compliance team, the uncomfortable possibility is that AI increases demand for expert review while removing some of the work through which the next generation of reviewers would have learnt.

What Anthropic counted as work

Anthropic combines internal employment and usage records, output measures, surveys and 17 employee interviews. It reports a declining share of junior technical hiring alongside increased agent use and greater emphasis on autonomy and judgement in hiring decisions. By August 2026, employee-directed interactive agent runtime was six times estimated employee hours.

Agent runtime includes inference and active tool use, pools subagents and excludes waits longer than 30 minutes. Human hours are estimated from headcount using a 40-hour week. The ratio therefore cannot be read as six times the useful work or a sixfold productivity improvement. Anthropic also describes employees facing a growing burden of reading and directing agent output.

The findings do not establish that AI alone caused the hiring shift. Company growth, changing work and hiring composition complicate the interpretation. A lower junior share is also different from a claim that all junior work or absolute technical employment disappeared.

For a marketing leader, faster drafting can move the constraint to review. A team may produce many more campaign variants while still depending on the same specialists to establish whether the proposition is accurate, appropriate and supported.

Coordination and review can absorb a marketing team’s AI productivity gains. Learning presents a less visible cost: the saving on a first draft is easy to count, while the loss of practice may take much longer to become apparent.

Which parts of routine work teach judgement?

The Alan Turing Institute’s CETaS briefing on professional skills, published on 9 September 2026, draws on legal and health-sector evidence to examine implications for intelligence analysis. It distinguishes the erosion of existing skills from “never-skilling”, where foundational capability is not developed because AI absorbs formative tasks. It is a cross-sector evidence synthesis, not an experiment demonstrating a particular training remedy.

An experienced reviewer relying heavily on a tool may lose practice. A new colleague receiving only finished analyses may never learn how to construct one. Both can become dependent on the system, but refresher practice will not address a skill that was never acquired.

A junior marketer preparing a mortgage campaign might learn most from comparing product terms and discovering why an attractive claim needs qualification. Manually copying the approved wording into a document may teach much less. Treating both as production tasks to automate would conceal that difference.

A junior colleague could first assess the product proposition, then use AI to examine it before discussing disagreements with an experienced reviewer. Research and drafting could still be faster, while the learner would have a judgement of their own to defend or revise.

Selected cases might provide enough exposure to difficult distinctions without requiring every piece of content to follow that route. The choice depends on what people will need to judge later; there is little reason to preserve manual copying simply because it once occupied a junior role.

Assistance could make some lessons easier to learn

AI could also help a junior colleague explore an unfamiliar requirement or ask questions they might hesitate to raise in a busy team. Reliable answers and encouragement to examine them could make learning easier. The risk is a polished assessment that offers little reason to inspect its sources.

Comparing an AI recommendation with the underlying evidence gives the learner something to do beyond accepting or rejecting the answer. They can identify what supports it, where an assumption enters and which missing fact would change their view.

Senior feedback matters here. Simply exposing staff to a large archive of approvals does not explain why a decision was appropriate. Product conditions, audiences and regulatory interpretations may have changed. The conditions behind an expert decision make it useful both as organisational memory and as material for an apprenticeship.

A decision record can show why an exception was accepted and what uncertainty remained. Discussing whether that reasoning applies to a new case would make the record useful for development as well as reference. Copying the conclusion would teach much less.

Experts need time to teach as well as review

An organisation adopting AI often asks how much work it can remove from a role. It should also examine what work remains and whether people are equipped to perform it. Reviewing a large volume of machine output may require different skills from producing a smaller volume personally, particularly when errors are plausible and inconsistent.

Selected independent exercises could help a compliance team see whether staff can recognise missing evidence without the tool. Case discussions could reveal whether they understand a correction or have simply accepted it. These are suggestions to test, not training methods established as effective by the Anthropic study.

Development also requires protected attention from experts. If the productivity plan fills every saved hour with additional production, there may be less capacity for the discussion through which judgement is transferred. The business can become dependent on a shrinking group of specialists while congratulating itself on faster output.

Junior roles will have good reasons to change, and AI may let people attempt more sophisticated work earlier. What needs deliberate replacement is the learning lost when a task disappears. A team that saves its experts time on production but leaves them no time to teach may eventually struggle to find anyone ready to take their place.