Why one AI exec says token counts are a trap

The warning comes as corporate AI budgets climb. Gartner expects global spending on AI software to top $100 billion this year. CFOs want return-on-investment numbers. Token counts provide a simple number, one that can be gamed.

A worker can feed a long document to a model and inflate the token figure without extracting real value. A better metric is whether the AI cut the time to draft a report or improved its accuracy, Gonzalez said.

Synthesia is not alone in rethinking the approach. Some large enterprises now separate usage tracking from impact tracking. They look at features adopted, error rates, and user satisfaction scores.

Gonzalez acknowledged that tracking outcomes is harder than tracking tokens. It requires clear baselines and consistent follow-up. She argued that the effort pays off when managers can point to genuine productivity gains.

The company plans to share more details on its measurement framework later this year. No date has been set.