Does every feature we build with AI need a token budget?

2026-06-05 · Source: Thoughtworks Insights · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Project & Product Management, Corporate Strategy & Leadership · Depth: Intermediate, short

Summary

Uber's CTO Praveen Neppalli Naga exhausted the company's 2026 AI budget by April, primarily due to rapid adoption of Claude Code, which saw usage jump from 32% to 84% among 5,000 engineers in four months, incurring \$500-\$2,000 monthly API costs per heavy user. This "tokenmaxxing" trend is mirrored at Meta, where employees consumed 60.2 trillion tokens in 30 days, costing approximately \$900 million. The article advocates for a hypothesis-driven approach to AI feature development, requiring a visible, comprehensive token budget for each feature. This budget should encompass build, run, and critically, maintenance costs, which include model deprecation and re-evaluation. The shift to AI accelerates development but moves the bottleneck to understanding product impact, making careful budgeting essential to avoid financially risky, high-cost features.

Key takeaway

For AI Product Managers evaluating new features, you must implement a rigorous token budgeting process covering build, run, and long-term maintenance costs. Tie budget allocation to clear business hypotheses and validate them cheaply. This approach prevents "tokenmaxxing" and ensures your team invests only in features where the run budget aligns with plausible business value, avoiding costly, unneeded capabilities. Implement circuit breakers to mitigate runaway agent expenses.

Key insights

AI features require comprehensive token budgeting across build, run, and maintenance, guided by hypotheses.

Principles

• Budget AI features across their full lifecycle.
• Hypothesis-driven development reduces financial risk.
• Underestimate AI feature maintenance at your peril.

Method

Estimate value and set a token budget for each AI feature, covering build, run, and maintenance, tied to a stated hypothesis for validation.

In practice

• Attach token budgets to every ticket.
• Track build, run, and maintenance budgets.
• Implement circuit breakers for agents.

Topics

• AI Budgeting
• Token Costs
• Large Language Models
• Hypothesis-Driven Development
• Feature Lifecycle Management
• Circuit Breakers

Best for: CTO, VP of Engineering/Data, Executive, Director of AI/ML, AI Product Manager, Consultant

Related on AIssential

• 📈 Data to start your week: The cost of tokenmaxxing — Exponential token consumption is rapidly increasing AI operational costs, challenging traditional budgeting and cost management strategies.
• Your LLM bill is not your infra bill: a budgeting catalog for AI-feature SaaS — AI costs are a metered utility, not fixed infrastructure, requiring proactive behavioral budgeting to prevent unexpected spikes.
• Intelligence Per Dollar — AI model evaluation is shifting from raw performance to cost-efficiency, prioritizing "intelligence per dollar."
• Perplexity - Perplexity — Unrestricted AI access without cost controls leads to "tokenmaxxing" and massive, unmanaged spending.
• AI Costs Are Outpacing Marketing Budgets, So How Do You Strategize? — Escalating AI costs, fueled by agentic AI's high token consumption, demand strategic management focused on value and task-appropriate model selection.
• Executive Briefing: Uber Burned Its Entire AI Budget Early. The Bill Was Trying to Tell Them Something. — High token demand, not just cost, signals useful AI adoption, challenging simplistic ROI metrics for agentic work.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Thoughtworks Insights.