Cut your AI bill by teaching staff which model to use and capping daily use
Experts recommend training and daily limits to contain costs, while AI companies say rapid change makes precise forecasting hard.
WSJ What's News · Why Businesses Can’t Budget for AIStephanie Stam, a visual editor who covers artificial intelligence, asked experts what a business owner should do as AI bills become unpredictable. Their answer was habits, not forecasts: teach staff to match each job to a suitable system, and set a daily allowance for AI use for each employee.
A model is simply a version of an AI system, with some built for quick questions and others built for harder work like computer programming. A token is the unit of work the business pays for, with every choice, check, and retry adding to the count.
Stam showed why the match matters with the same request to analyze a frame from a YouTube video. The cheaper Gemini 3 Flash failed after 952 steps and cost $14, while the stronger Gemini 3.1 Pro solved it in 85 steps for about $1.
Which model and effort level an employee picks matters because each choice changes how many tokens are used.
Then add a guardrail. Set a daily allowance per employee, and track who is approaching it. Stam's test showed how runaway use happens: an agent can keep retrying and looking elsewhere for answers, burning tokens without finishing the task. The allowance sets a limit that can stop that kind of retrying before spending climbs further.
Any list of which system to use will need updating, because AI companies say the technology keeps evolving and precise forecasting is hard.
As a suggestion for putting this into practice, rather than advice quoted from the experts, a company could start with one team and one rule: no complex work on the cheap, fast model without checking first. Name the two systems they should use, and when to switch.
Questions people ask
What exactly is a token?
A token is a small chunk of text or data that an AI system reads or writes, roughly a few characters or part of a word. Providers count those chunks to measure work and set prices, so longer thinking, retries, and tool use create more tokens. A short answer may cost a few hundred tokens, while an agent working for many steps can run into the hundreds of thousands.
How do I know which model is right for which task?
Use lighter, faster, cheaper models for simple drafting, summaries, and routine questions. Use stronger, slower, more expensive models for coding, math, legal or financial analysis, and multi-step research. If a task needs several steps, files, or outside lookups, start with the stronger model, since a weak model may take many more steps and still fail.
What should a daily token limit be?
There is no universal number, because prices and work vary widely by company and provider. Start by checking your admin dashboard for average tokens per employee per day, then set an initial cap somewhat above that and tighten it over a few weeks. Pair the cap with alerts, approvals for exceptions, and a review of the highest-use cases.