Beyond tokenmaxxing: Is it time for enterprises to scrutinise AI returns | Artificial Intelligence News

The race to adopt artificial intelligence (AI) across workplaces is entering a new phase, with technology companies now questioning whether higher AI usage necessarily translates into greater productivity. Amazon has cautioned its software engineers against “tokenmaxxing”, a practice of maximising AI consumption to appear more productive.

 

While Amazon is the latest to issue such a warning, calls to rationalise AI usage in the workplace have been made before.

 

In August, US-based technology Microsoft had reportedly cautioned employees against treating token consumption as a measure of productivity. According to reports by 404 Media and ITPro, the company had issued internal guidance asking engineers to focus on business and customer outcomes rather than maximising AI usage.

  

Likewise, social media entity Meta in July said it is considering token budgets for engineers. According to media reports, Instagram head Adam Mosseri said the cost of running AI tools could become comparable to an engineer’s employment cost, making it necessary for companies to allocate token budgets in the same way they manage computing resources and payroll.

 

Cab-aggregator service provider Uber also questions the returns from AI spending. In May, the company’s president and chief operating officer, Andrew Macdonald, said it was difficult to establish a direct link between higher AI consumption and the number of useful features delivered to customers.

 

The company had exhausted its annual AI coding budget by April, according to earlier reports citing its technology leadership.

 

From Meta’s discussions about employee token budgets to Uber’s spending limits, companies are looking for ways to measure the returns from AI investments rather than usage alone.

What is tokenmaxxing?

Tokenmaxxing refers to the practice of maximising the number of AI tokens consumed to demonstrate frequent or extensive use of AI tools. The term gained attention after employees at technology companies began competing to show how much AI they were using, with token consumption sometimes treated as a measure of productivity or technical ability.

 

How do AI models process information in tokens

 

Tokens are the basic units of text and other information that AI models process. They can represent whole words, parts of words, punctuation or other elements, depending on the language and the model’s tokenisation system.

 

When a user enters a prompt, an AI model breaks the information into tokens, processes it and generates a response, which is also measured in tokens. For example, a short question may require only a few dozen tokens, while a lengthy document, detailed instruction or coding task can involve thousands.

 

AI usage generally involves two broad categories of tokens:

• Input tokens: The information provided to a model, including user prompts, instructions, documents and relevant conversation history.
• Output tokens: The information generated by the model in response, including explanations, summaries, code and other content.

The amount of information processed can increase when users provide lengthy instructions, upload large documents or ask models to retain and refer to extensive conversation histories. 

Why does higher AI usage increase costs?

Many AI services offered to businesses through application programming interfaces (APIs) follow usage-based pricing. Companies are charged according to the number of input and output tokens processed, with rates varying across models and service providers.

 

Longer prompts and larger context windows can increase the number of input tokens. Repeatedly asking a model to revise an answer, analysing the same information across multiple requests or running unnecessary agent workflows can add to consumption.

 

Autonomous agents can make this more complicated. An agent assigned a task may repeatedly call a model, access external tools, review the results and attempt another approach if its first response is unsuccessful. In complex workflows, these repeated interactions can increase inference costs even when the final output is relatively short.

 

The cost can also rise when employees use more advanced models for tasks that could be handled by smaller or less expensive ones.

 Even as the price of processing individual tokens declines, overall spending can increase if companies significantly expand the volume of AI usage. The challenge for businesses, therefore, is to ensure that additional consumption produces enough productivity gains to justify the expense. 

What does this mean for enterprise AI adoption?

The shift towards cost controls does not mean technology companies are slowing AI adoption. Instead, businesses are moving beyond tracking how often employees use AI to measure its impact on actual work. For software teams, this could mean assessing whether AI improves code quality, reduces development time or speeds up product releases. Other departments may track improvements in customer service and operational efficiency.

 

Companies may also allocate AI budgets based on team requirements, encourage the use of smaller models for routine tasks and reserve more expensive models for complex work. Monitoring token consumption could help businesses manage costs without restricting AI use.

 

Amazon’s guidance reflects this shift. While AI will remain central to software engineering, employees are expected to assess where it adds value. As AI agents and coding assistants become more common, companies will need to balance experimentation with financial discipline, focusing on useful work delivered rather than tokens consumed.