Key facts
- Atlassian has implemented monthly AI spending limits for its research and development staff.
- The AI "wallets" have caps ranging from $500 to $2,000 per employee.
- This policy contrasts with the "tokenmaxxing" trend where other tech companies encouraged maximum AI usage.
- The rising costs are attributed to AI agents autonomously undertaking tasks.
- Companies are considering using less powerful models and open-source AI to manage expenses.
Software firm Atlassian has introduced monthly spending limits for its employees' use of artificial intelligence tools, capping individual "wallets" at a maximum of $2,000. This move comes as other technology companies have seen AI costs escalate, with some encouraging employees to use as much AI as possible, a practice known as "tokenmaxxing."
Atlassian's new policy, implemented for its research and development team, aims to manage the rising expenses associated with AI, particularly driven by autonomous AI agents that initiate tasks and generate numerous prompts. While Atlassian has always provided AI budgets, the introduction of these capped wallets signifies a shift towards more controlled usage.
Companies like Uber have reportedly exceeded their AI budgets quickly, and Amazon has advised employees to cease using AI solely for the sake of usage. The cost of AI tokens, which measure AI responses, can quickly accumulate. OpenAI charges $5 per 1 million tokens for its GPT-5.6 Sol model, while Anthropic's Claude models cost $10 per 1 million tokens.
A survey by PureProfile found that 80% of senior Australian staff at AI-using companies are concerned that high AI usage is being mistaken for productivity gains, with 32% having scaled back AI deployments due to cost. Experts suggest that implementing spending caps is a "smart" approach to incentivize appropriate AI behavior and prevent inefficient use.
To further drive down costs, companies are exploring strategies such as utilizing less powerful AI models for simpler tasks and investigating open-weight or open-source models that can be run on internal systems.