When Companies Mandate AI-Only Coding, Budgets Override Real Productivity

A junior engineer at a Norwegian tech company of about 200 employees shared a Reddit post describing three phases of the firm’s AI‑coding policy. In April, senior leadership declared traditional IDEs outdated and instructed engineers to avoid code editors, using AI in the terminal for virtually all coding tasks.

By the end of May, the company introduced a hard token budget: junior engineers received roughly $500 per month and senior engineers $1,500, while the most powerful models were replaced with more basic versions.

At the end of June, a new rule tied unused token budget to bonuses—half of any unspent allocation would be added to salaries. This incentive caused almost all developers, especially seniors, to dramatically cut back or stop using AI, reverting to pre‑2023 hand‑written code practices.

The engineer worries that, as a newcomer, he may be scapegoated if management questions the policy, leaving him torn between following colleagues or adhering to the original “AI‑first” directive.

The post argues that when a coarse metric (token consumption) is used to reward behavior, people optimize for the metric rather than the underlying goal. This mirrors a broader trend observed in many large tech firms: early in the year, companies gamified AI usage with leaderboards and token‑based competitions, sometimes consuming tens of thousands of billions of tokens and costing hundreds of millions of dollars.

Examples include Meta’s internal leaderboard tracking tens of thousands of employees, Uber’s AI budget for 2026 being exhausted in four months, and similar token‑capping policies at Amazon, Walmart, Microsoft, Tesla, Alibaba, Baidu, and ByteDance.

The shift is attributed to budgets being set on outdated assumptions while AI consumption rates change rapidly. Tools like Claude Code, Codex, and autonomous agents dramatically increase token usage, making earlier forecasts inaccurate.

Rewarding token volume rather than actual output creates a disconnect: high token use does not guarantee higher productivity, and employees may appear productive without delivering real value.

The author questions whether the goal is to “distill” or replace employees, or genuinely integrate AI. He notes that the promised ten‑fold efficiency gains rarely materialize; instead, organizations incur high costs without corresponding performance improvements.

Potential reasons for the backlash include compliance risks, data privacy concerns, the steep cost of top‑tier models, and diminishing AI performance.

Conversely, the imposed limits also signal that the tools have value—companies set caps instead of banning AI outright because employees are willing to spend significant money on them.

Ultimately, the author concludes that before financials are clarified, any proclaimed AI embrace remains superficial; the transition from a loose, growth‑driven model to a more disciplined, budget‑controlled approach reflects the true state of AI adoption in software development.

“If that engineer earning $500,000 per year consumes less than $250,000 worth of tokens, I would be very concerned.” – Jensen Huang, NVIDIA