While evals involve testing, their purpose isn't just to report bugs (information), like traditional QA. For an AI PM, evals are a core tool to actively shape and improve the product's behavior and performance (transformation) by iteratively refining prompts, models, and orchestration layers.

Treating AI evaluation like a final exam is a mistake. For critical enterprise systems, evaluations should be embedded at every step of an agent's workflow (e.g., after planning, before action). This is akin to unit testing in classic software development and is essential for building trustworthy, production-ready agents.

While the goal is autonomous improvement, deploying these systems safely in production requires human oversight. Implement mandatory human-in-the-loop steps, specifically code reviews for any proposed changes to the agent or its evaluation logic, before shipping to users.

When an AI tool makes a mistake, treat it as a learning opportunity for the system. Ask the AI to reflect on why it failed, such as a flaw in its system prompt or tooling. Then, update the underlying documentation and prompts to prevent that specific class of error from happening again in the future.

Do not blindly trust an LLM's evaluation scores. The biggest mistake is showing stakeholders metrics that don't match their perception of product quality. To build trust, first hand-label a sample of data with binary outcomes (good/bad), then compare the LLM judge's scores against these human labels to ensure agreement before deploying the eval.

Treating AI evaluation as a single, pre-launch check is a mistake. Model behavior drifts due to fine-tuning, infrastructure changes, and shifts in user queries. Production AI systems demand a continuous evaluation pipeline integrated into the deployment lifecycle to catch regressions and ensure ongoing reliability.

The prompts for your "LLM as a judge" evals function as a new form of PRD. They explicitly define the desired behavior, edge cases, and quality standards for your AI agent. Unlike static PRDs, these are living documents, derived from real user data and are constantly, automatically testing if the product meets its requirements.

Don't just assume a new AI workflow is better. Treat internal process changes with the same rigor as product features. Apply a hypothesis-driven framework to how your team operates, experimenting with new AI tools and methods, and validating whether they actually improve outcomes before committing to them.

Don't just automate tasks; automate quality control. Create an agent that reviews a core part of your app daily, grades it against a rubric you define, and automatically spins up a new "child" agent to fix anything that scores below a certain threshold, creating a virtuous cycle of improvement.