AI Changed What "Done" Means

“The Definition of Done is the team's commitment to quality. Without it, you are merely building up technical debt.” — Ken Schwaber, co-creator of Scrum and signatory of the Agile Manifesto.

For decades, the Definition of Done (DoD) has been one of the clearest commitments an Agile team can make to quality expectations. It creates shared confidence among teams, stakeholders, and customers by making explicit what must be true before work is considered complete and ready for delivery.

Most Definitions of Done, however, were designed for a world where humans authored the code, created the tests, and validated the resulting artifacts. That assumption is no longer guaranteed.

AI has moved well beyond autocomplete. It now contributes across increasingly complex parts of the software delivery lifecycle—from code and tests to documentation, infrastructure, and even architectural decisions.

This shift exposes a critical blind spot in modern engineering: What evidence do we need before we can trust what AI helps us build?

Recent analyses highlight a growing risk in AI-assisted delivery: AI can produce plausible but incorrect results and take unintended actions with real operational consequences. Tools and platforms—from coding assistants to agentic systems like Timothy Assist, Orbit, and FRIDA—raise the stakes for verification.

Different incidents. Different systems. The same underlying lesson: AI capabilities are advancing faster than some of the practices organizations use to verify, govern, and build confidence in their outcomes. The risk changes further as AI moves from assisting human decisions to taking actions through tools and production systems.

This is not an argument against AI. It is an argument for modernizing our engineering standards to match the realities of AI-assisted delivery.

Passing automated tests, completing code reviews, and satisfying acceptance criteria remain essential engineering practices. But AI-assisted delivery raises questions those practices alone may not fully address.

This is where the emerging concept of an AI-Aware Definition of Done becomes valuable. It extends the Definition of Done to explicitly address the realities of Human–AI collaboration, making accountability, verification, transparency, traceability, and risk-based validation explicit.

When the way we build changes, the evidence behind “Done” must change with it. Quality is not assumed; it is demonstrated.

When organizations encounter new risks, the instinct is often to add another governance layer: a committee, approval gate, or another compliance checklist. Instead, strengthen the practices teams already use rather than creating another process.

Most Agile teams already maintain a Team Working Agreement that defines how they collaborate, make decisions, and deliver value. Do not replace it. Extend it.

An AI Working Agreement (AI-WA) establishes clear expectations for Human–AI collaboration, covering accountability, security boundaries, approved tools, verification, and traceability of AI-assisted work.

At the center of the AI-WA should be the AI-Aware Definition of Done, making one principle explicit: If AI contributes to delivery, the work must still meet the team's standard for quality.

Organizations will not realize AI's productivity gains if every production incident erodes confidence in AI-assisted delivery. The leaders of the AI era will not simply be those who adopt the most powerful AI capabilities, but those who pair innovation with engineering discipline.

AI changed what must be true for work to be complete, but it didn’t change the purpose of the Definition of Done. It raised the bar for the evidence we need before we can confidently say, “We’re done.”

Adopting an AI-Aware Definition of Done entails a fundamental shift in engineering culture. At MK Timothy & Company, delivery stays tied to real operators and verifiable outcomes — so teams can move faster without lowering the bar on quality.