Large employers that encouraged workers to consume as many AI tokens as possible are reassessing the practice as bills rise without comparable gains in output. Moody’s AI analytics head Vincent Gusdorf warned that generative systems make it easy to produce work an organization does not need. Bain consultant Jue Wang said token costs at some large clients had been doubling nearly every other month.
A hypothetical $200 monthly cost across 20,000 developers would equal $4 million each month. Companies are adopting model routing so simpler requests go to less expensive systems. More capable and costly models can then be reserved for work that requires them.
Technology executives had briefly promoted high token use as a sign of ambitious employee adoption. Consumption is an input measure and does not establish accuracy, time saved, revenue or customer benefit. Model routing requires organizations to classify tasks and decide what data may reach each provider. Cost analysis is complicated because a subscription price is only one component. Review time, errors, security controls and workflow changes can increase or decrease the total economic result.
Microsoft CEO Satya Nadella also raised concern about customers paying fees while sending proprietary data to providers. Lower-cost open models, including systems developed in China, are adding pricing pressure. Software organizations previously abandoned lines of code as a productivity proxy for similar reasons. The AP account drew on executives, consultants and workplace examples rather than one audited economy-wide study. It establishes a visible change in corporate practice without quantifying a universal return on AI spending.
The reporting available at the edition deadline did not resolve this point: no single audited dataset measured token costs and productivity across all industries. The next dated records are company disclosures of task-level AI returns and adoption of model-routing and lower-cost systems.