Seattle: Amazon has issued updated guidance for its software engineers on using artificial intelligence (AI) at work, emphasising customer outcomes, human judgment and measurable productivity rather than the volume of AI-generated code or tokens consumed.
The internal memo, circulated to employees and subsequently reported publicly, outlines how the company expects Software Development Engineers (SDEs) to incorporate AI into their daily work. It encourages engineers to experiment with AI tools while maintaining established engineering standards and focusing on delivering useful products to customers.
The guidance also cautions against “tokenmaxxing”, a practice in which employees use AI tools excessively to inflate usage metrics without producing meaningful results. Amazon’s message is that adopting AI should improve software development and customer experiences, not simply make employees appear more productive.
The memo also places responsibility on engineering leaders to understand AI tools through practical experience and to assess whether their use leads to measurable improvements in software delivery.
Amazon promotes the idea of the Renaissance Developer
Amazon has described the emerging engineering role as that of a “Renaissance Developer” — a professional who combines AI capabilities with human judgment, technical knowledge and an understanding of the business.
In the past, Amazon’s engineering guidelines generally focused on enduring principles rather than requiring developers to adopt specific technologies. Major changes, including cloud computing, microservices and serverless architecture, altered how software was built without fundamentally changing the expectations placed on engineers.
The company believes AI is different because it is changing software development at a faster pace and affecting more parts of the engineering process.
According to the memo, engineers need to learn how AI tools work and understand where they can improve their daily tasks. Those who fail to develop the necessary skills risk falling behind as the technology becomes more integrated into software development.
However, the guidance does not suggest that AI usage alone defines an engineer’s performance.
Amazon continues to emphasise the importance of producing reliable, high-quality software that meets customer needs. Engineers are expected to combine the speed offered by AI with their own technical expertise and judgment.
The Renaissance Developer concept therefore centres on using AI as a capability that strengthens engineering work, rather than treating it as a replacement for the fundamentals of software development.
What is tokenmaxxing, and why is Amazon warning against it?
One of the most notable aspects of the memo is Amazon’s warning against “tokenmaxxing”.
The term refers to attempts to increase AI usage metrics by assigning unnecessary or low-value tasks to AI agents, thereby consuming more tokens without necessarily improving productivity.
Tokens are units of text processed by AI models. The number of tokens used can indicate how much activity takes place within an AI system, but it does not automatically measure the quality or usefulness of the resulting work.
For example, an engineer might generate large amounts of AI-assisted code or repeatedly run an agent on tasks that do not meaningfully advance a project. The resulting usage figures could look impressive even if the work does not lead to a better product or faster delivery.
Amazon’s guidance warns that such behaviour can create the wrong incentives.
The company has also raised concerns about using AI token consumption and the number of AI-assisted code changes as primary measures of employee performance. These figures can be misleading if they are not linked to actual outcomes.
The memo reflects Amazon’s broader emphasis on frugality, a principle that encourages employees to use resources carefully and avoid unnecessary expenditure.
Excessive AI usage can increase computing costs without delivering corresponding benefits. By discouraging tokenmaxxing, Amazon is seeking to prevent employees from treating usage statistics as a substitute for meaningful engineering output.
Amazon wants productivity measured through real results
Instead of measuring engineers by how much AI they use, Amazon wants performance assessments to focus on the results of their work.
The company is placing greater emphasis on long-term productivity, successful software releases and the speed with which useful features reach customers.
Marc Brooker, a vice-president at Amazon Web Services (AWS), told CNBC that the company was focusing on “feature velocity”. The concept measures how effectively teams deliver successful updates and new features to users.
This approach shifts attention from the amount of activity generated during development to the value ultimately delivered.
An engineer who uses AI to solve a complex problem and successfully releases a useful feature may create more value than someone who generates a much larger volume of code that requires extensive reworking.
Similarly, an AI tool that reduces repetitive work or helps engineers identify problems earlier could improve productivity even if its overall token consumption is relatively modest.
The memo encourages managers to assess whether AI adoption is actually improving delivery, rather than assuming that greater usage necessarily translates into better performance.
For Amazon, the important question is whether AI helps teams build and deliver better products.
Managers must examine both low and high AI adoption
Amazon’s guidance also calls on leaders to examine how different teams are using AI.
The memo highlights the need to pay attention to both the 10th percentile of teams that are adopting AI less effectively and the 90th percentile of heavy AI users.
The purpose is to understand why adoption differs across the organisation and whether greater usage is producing tangible improvements.
Teams using AI relatively little may need additional training, better access to tools or help identifying suitable applications. In other cases, limited adoption could reflect the nature of their work or the fact that available tools do not provide sufficient benefits for their particular tasks.
At the other end of the spectrum, teams using AI extensively must demonstrate that the additional activity leads to faster development, better-quality software or other measurable gains.
High usage alone is not proof of success.
By examining both groups, managers can identify effective practices, understand barriers to adoption and avoid rewarding activity that does not translate into business value.
The approach also recognises that AI adoption is unlikely to be identical across every engineering team.
Amazon expects leaders to use AI hands-on
The memo places responsibility on technical and organisational leaders to develop practical experience with AI tools.
Rather than relying solely on reports from employees or general claims about the technology, managers are expected to understand how AI performs in real production work.
They should be able to identify where a tool has helped, where it has fallen short and whether it has improved the team’s overall output.
Amazon’s guidance links this expectation to its leadership principle of “Learn and Be Curious”. Leaders are encouraged to make time for themselves and their teams to explore AI tools, experiment with different workflows and share what they learn.
Hands-on experience can help managers make more informed decisions about where AI should be introduced and which applications deserve further investment.
It can also prevent organisations from adopting tools simply because they are popular or because competing companies are using them.
The message is that leaders should understand both the capabilities and limitations of the technology before making decisions about its use.
Employees can decide when AI is not useful
Although Amazon is encouraging engineers to develop their AI skills, the memo does not call for the technology to be used in every task.
Instead, it acknowledges that some types of work may not benefit from the AI tools currently available.
The guidance tells employees that if a particular task does not improve through AI assistance, that conclusion should be treated as a finding rather than a failure.
Engineers are encouraged to document situations in which a tool does not help and share that feedback with their managers.
This is an important distinction because AI systems can introduce their own costs, including the time required to review generated code, correct errors, verify results and integrate outputs into existing systems.
For a simple or well-understood task, using an AI tool may add unnecessary complexity. In other situations, the tool may produce results that require so much correction that the expected productivity gains disappear.
Amazon’s approach allows engineers to exercise judgment instead of using AI merely to satisfy an adoption target.
The emphasis remains on selecting the method that produces the best outcome for the customer and the business.
Customer understanding remains central to engineering
Another major theme in the memo is the relationship between engineers and the customers who use their products.
Amazon has observed that employees who understand their products and businesses closely are often better positioned to identify useful applications for AI.
As organisations grow, engineers can sometimes become distant from the customers their work is intended to serve. The company wants engineers at all levels to maintain a strong understanding of product direction and customer requirements.
That knowledge can help developers determine where AI is likely to make a meaningful difference.
An engineer who understands a customer’s problem may use AI to speed up development, automate repetitive tasks or explore alternative solutions. The same engineer may also recognise when AI introduces risks or provides no practical advantage.
The ability to decide when not to use AI is therefore part of the judgment Amazon expects from its engineers.
The memo’s broader message is that technological experimentation should remain connected to the purpose of the product. AI is valuable when it helps solve real problems, not simply when it is used more frequently.
What Amazon’s AI rulebook means for software engineers
Amazon’s updated guidance reflects a shift in how companies are thinking about AI adoption in software development.
Early discussions around AI productivity often focused on how much code developers could generate or how many tasks AI agents could complete. Amazon’s memo instead places greater emphasis on the quality of the output, the efficiency of delivery and the value created for customers.
For engineers, this means learning to use AI tools effectively while retaining responsibility for the code and products they deliver.
For managers, it means evaluating AI adoption through measurable results rather than relying on token consumption, code volume or other activity-based metrics.
The guidance also recognises that different teams may require different approaches. Some workflows could benefit substantially from AI, while others may see limited improvements until the tools become more capable.
Ultimately, Amazon’s message is not to use AI as much as possible. It is to understand the technology, apply it where it helps and remain accountable for the final result.