As artificial intelligence (AI) use ramps up by many companies, some are reportedly facing sticker shock as they run out of AI “tokens,” which can lead to extra costs.
A report from Bloomberg last week said that Uber was capping AI use by employees after burning through its entire 2026 AI coding budget in just four months, and that’s part of a growing number of U.S. media reports on companies being caught by surprise by costs from using up AI tokens faster than expected.
AI tokens are used in Large Language Models (LLMs) like ChatGPT and Anthropic’s Claude and act as a sort of meter that AI providers use to charge businesses based on consumption of the tool’s capabilities.
The result can mean surprise costs for businesses that can quickly add up.
“We see a lot of these mistakes happening. So they (businesses) buy the licence, they run the pilot, but they’re not really changing their behaviour,” says Patrick Farrar, CEO of AI for Canadians and the head of AI at DMZ, a technology incubator and startup ecosystem at Toronto Metropolitan University.
“They’re not doing workflow changes within the organization and they’re not training people within the organization. That’s where you’re starting to see, really, these costs add up.”
This also comes as Ottawa’s recently announced national AI strategy aims to boost business growth in Canada, including through the implementation of AI in business workflows and educating workers and students.
So how exactly do AI tokens work?
Related Videos
Most consumers and businesses are able to access several free versions of AI models that feature scaled-down or limited capabilities. For companies, though, many need enterprise licenses to use AI models in commercial work and can get a set number of tokens through their contracts or pay as they go.
If businesses or consumers run out of tokens, then they get charged for each additional token spent.
Get daily National news
Get daily Canada news delivered to your inbox so you’ll never miss the day’s top stories.
Farrar says that a typical prompt on an AI model like ChatGPT that spans a sentence or two and generates a short response may use a fraction of a token per word, while a longer conversation might use 1,000 tokens.
This, he says, can cost half a cent to two cents, while more complicated requests that involve visualizations and designs can be dramatically more expensive.
“Now you’re seeing businesses actually run real (AI) workflows every day. Even taking it a step further, you’re seeing companies use these agentic tools as they’re called now, so these tools that run autonomously,” says Farrar.
“And so those sort of agents, as they’re called, or these sort of autonomous tools can now run all day or every hour or all week and all month. And so that can burn upwards of 50 to hundreds of times more tokens in a single sort of conversation even. So that’s why maybe finance now is starting to see the bill on that.”
Canadian businesses are trying to innovate with AI where possible while trying to stay efficient with their use of AI token budgets.
“We’ve been really asking for a lot of organizations to push the usage of AI, to start collaborating with the AI tools so they can at least increase their knowledge and be ready for some of the bigger changes coming, but with that came some pretty significant costs,” says Annie Veillet, a partner at PwC Canada with a focus on AI engineering and data engineering practice.
“Depending on the models that were selected and made available to their employees or end users, the understanding of the cost that was going to come at the end of the usage wasn’t necessarily well understood.”
Veillet says many businesses are using these AI tools to streamline tasks that are perceived as being tedious or that are relatively easy to do but are time-consuming.
When implemented well, these tools can work well for companies, but if done on a large-scale, those that aren’t keeping track of their AI tokens could quickly run out and get a big bill.
“Now there’s a lot more guardrails put around it. ‘For your day-to-day type of tasks, please use these models, which are a lot less costly,’” she says.
“‘For some more complex, deep work, you can still, and we encourage you to work with some of the large language models that are a bit more costly because we expect a larger return on investment.”
© 2026 Global News, a division of Corus Entertainment Inc.
