Organizations are encouraging more employees to build with artificial intelligence, but relying on a single general-purpose tool for every task can create disappointing results and undermine confidence in the technology. When this happens at scale, workers may repeatedly refine prompts until they obtain something usable, a pattern described as “tokenmaxxing.”

The source argues that tokenmaxxing is less a failure of employees than a symptom of unclear objectives, insufficient guardrails and limited ways to measure AI spending and outcomes. Simply blaming staff or adding broad responsible-use training does not address those underlying problems.

Traditional procurement processes can also conflict with the speed of AI development. While approving enterprise software may take months, an employee can access a public chatbot and generate a landing page within seconds. If approved tools are too slow or restrictive, workers may turn to unvetted services, potentially entering proprietary information into public models and creating data-governance risks.

The resulting prototypes can introduce further problems. They may contain poor code, fail to follow corporate design systems or accessibility requirements such as WCAG 2.2, and omit basic authentication practices. Restricting access does not necessarily stop this behavior; employees may instead pay for subscriptions themselves, moving costs outside the organization’s immediate financial visibility.

The article also links unmanaged AI use to pressure for faster returns. A Teneo survey of more than 350 public-company CEOs and 400 institutional investors found that 53% of investors expected positive returns from new AI initiatives within six months or less, compared with 16% of large-cap CEOs who believed they could meet that timeframe. The CEOs also reported that fewer than half of their current AI projects were generating a positive return on investment.

At the same time, token use is expected to grow significantly. Goldman Sachs Research forecasts a 24-fold increase in monthly token consumption, reaching 120 quadrillion tokens between 2026 and 2030, while semiconductor providers are delivering annual reductions of 60% to 70% in the cost per token. Lower prices could encourage greater usage, particularly as organizations deploy autonomous AI agents.

Agents can consume substantially more tokens than a human working through a prompt because they repeat queries in sequences. An agent assigned an unclear task may continue operating long after a person would have questioned whether the work was worthwhile. Some of that spending goes toward prototypes that never reach production.

Gartner found that at least half of Generative AI projects had been abandoned after the proof-of-concept stage by the end of 2025. The reasons included poor data quality, inadequate risk controls, rising costs and unclear business value. Prototypes that require engineers to rebuild them from the beginning can also consume both token budgets and valuable engineering time.

This can deepen distrust among technical teams. Stack Overflow’s 2025 developer survey found that 46% of developers distrusted the accuracy of AI output. “Almost right” answers were the main frustration for 66% of developers, while 45% said debugging AI-generated code took more time.

The proposed alternative is not to limit AI access, but to make its use more specific and accountable. Giving non-engineering teams the ability to build can help create software shaped by people who understand the business requirement. However, each use should have a clearly defined purpose, and employees should be able to explain what a particular token expenditure was intended to achieve.

Instead of treating a general-purpose chat box as the default interface, organizations could provide workflows and agents tailored to specific roles. These environments could include execution limits, brand guidelines, approved fonts, security requirements and pre-built components. Embedding those controls would help employees produce applications that are closer to compliant and production-ready without requiring separate requests for every safeguard.

The source describes one mid-sized company where a three-person data analytics team built an agent to analyze information across its systems. The implementation saved each team member between 12 and 13 hours of manual work per week. The agent had a defined task and operated within predetermined parameters rather than functioning as an unrestricted chatbot.

The central argument is that falling token prices do not remove the need for financial oversight. Organizations should first define, in straightforward terms, what they want AI to accomplish and then align tools, limits and spending with that purpose. Without that clarity, an unexpected bill is only a symptom of a broader failure to establish what role AI should play.

Sources: TechRadar.