AI spend without capability spend is a productivity bet, not a productivity plan

An AI rollout can look financially disciplined on a spreadsheet and still shift work back to employees. The licence sits in technology spend. Time spent checking unreliable output, learning escalation rules or recovering from mistakes often lands elsewhere. Cutting a training budget does not make that work disappear.

Michael Edmondson?s November 5 Yahoo Finance commentary draws on different sources. SHRM?s 2026 benchmarking found median L&D spending per employee down 28% from 2025. A Careerminds survey, as reported in the commentary, found 47% of participating organizations increasing AI spending while reducing learning budgets. These figures describe spending choices. They do not show whether workers became more capable, or whether AI improved the work. The two measures also come from different samples.

I would examine one workflow before deciding whether development funding is too high or low. In a service team, identify which customer interactions AI now drafts, who checks accuracy, who handles exceptions and who takes responsibility for incorrect answers. If supervisors must review more work, a cost may have moved from the learning budget into management time. That matters to performance even when tool adoption is high.

The investment decision should name the task that must improve, the standard for acceptable output, and the support people need to reach it. Some teams may need practice with difficult exceptions. Others may need a better review screen, clearer authority to override the tool, or time for supervisors to coach. A generic AI course would miss those conditions. A blanket cut in learning spend could miss them too.

Finance and operations should examine rework, cycle time and error rates alongside licence fees and development spending. If verified output improves, lower formal training expenditure may be reasonable. If defects rise or managers absorb more review work, the apparent saving is incomplete. The cited spending surveys cannot settle that question. Evidence from the changed workflow can.