Amazon Warns Engineers Against Tokenmaxxing as AI Costs Mount
Amazon’s warning is less about employee behavior than about a maturing operating model for generative AI in engineering. Once AI assistants become a shared production dependency, token usage stops being a novelty metric and starts looking like any other metered infrastructure cost: something that needs governance, observability and outcome-based accountability.
For engineering leaders, the practical question is not whether teams use AI heavily or lightly, but where AI produces a better delivery profile. That means correlating AI use with indicators such as defect escape rates, review latency, lead time, rework, incident volume and customer-facing feature throughput. Without that link, token counts can reward expensive activity while hiding degraded code quality, architectural drift or unnecessary generated changes that increase maintenance burden.
This also has a platform-engineering implication. Organizations adopting coding agents and agentic workflows will need FinOps-style controls for AI consumption: budget guardrails, per-team visibility, workload classification and policy around high-cost tasks such as repeated code generation, autonomous retries and large-context operations. Otherwise, AI spend can scale faster than the engineering value it creates.
There is also an architectural nuance behind Amazon’s “Renaissance Developer” framing. Teams still need engineers who know when not to invoke AI, which tasks require stronger human judgment and where domain context matters more than generation speed. The most effective AI adopters are unlikely to be the teams with the highest usage; they will be the ones that treat AI as a targeted engineering capability, not a proxy for productivity.
Amazon has issued new guidelines for its software engineers that discourage excessive use of AI tools simply to boost productivity metrics, signaling a shift in how the company evaluates AI’s contribution to software development.
The guidelines address tokenmaxxing, a practice in which employees use huge volumes of AI tokens to demonstrate their use of the technology. Amazon warns that rewarding this behavior can raise computing costs without producing corresponding improvements in software development.
In a memo to Amazon engineers, Terk Terkowitz, the company’s VP and Distinguished Engineer, Worldwide Amazon Stores, outlined revised expectations for software developers. The guidelines emphasize customer results, software quality and business performance rather than the volume of AI-generated work.
Amazon’s engineering leadership, including Senior Principal and Distinguished Engineers, developed the guidelines based on observations of successful software developers. The company, to be sure, isn’t reducing its commitment to AI. Instead, it wants managers to distinguish between productive AI adoption and activity that merely looks productive.
The new policy is part of an overall effort among companies to determine to what extent AI tools deliver measurable business value, and if they do, which strategies are the most productive.
Measuring AI Productivity
A big issue here is the reliability of AI productivity measurements. Common metrics include token consumption, adoption rates, code modifications and completed code reviews. While these figures can be useful, they can also create incentives that work against efficiency. An engineer who generates more code, for example, hasn’t necessarily produced better software.
Amazon recommends evaluating AI adoption at both ends of the adoption spectrum. Managers should examine teams in the 10th percentile of AI adoption to understand why they are making limited use of the technology. They should also study teams in the 90th percentile to determine whether extensive AI use produces better results.
This approach acknowledges that AI usage data has value, but requires additional evidence to establish productivity gains. Amazon also recognizes that some engineering tasks receive little benefit from current AI tools. Developers are encouraged to identify these limitations and explain them to managers rather than use AI unnecessarily.
Perhaps the most compelling point here is one that developers already know: usage statistics alone offer limited evidence of a return on investment.
AI Spending Raises Questions
Amazon’s revised policy follows reports that some employees had manipulated internal AI usage measurements. In May, employees reportedly used MeshClaw, an internal AI agent that can deploy code and interact with Slack, to raise their token counts. Amazon subsequently discontinued an internal usage leaderboard.
Other major companies face similar concerns about AI spending. Uber president and COO Andrew Macdonald acknowledged in May that greater token consumption had not yet demonstrated a clear connection to valuable consumer features. Uber had reportedly exhausted its 2026 Claude Code budget by April.
Accenture is also examining how enterprises can control AI costs as organizations adopt agentic workflows, which can require substantially more computing resources than conventional chatbots. The consulting firm is reportedly developing Token IQ, a tool intended to help customers understand and manage token spending.
Beyond cost management, Amazon’s guidelines describe a changing role for software developers. The company uses the term Renaissance Developer to describe engineers who combine AI capabilities with technical expertise, customer knowledge and independent judgment.
Managers are expected to gain practical experience with AI tools, including understanding where the technology performs well and where it falls short. Amazon also wants developers to maintain stronger relationships with customers and understand the business needs behind their software projects.
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.