Our AI-augmented development workflow in practice

How AI fits the work

Our engineering team uses modern AI development tools as a default part of the workflow. The goal is not to replace careful engineering — it is to free senior people to spend more time on the parts of the job where judgment matters most.

In practice, this means AI is involved in many small moments throughout a working day, and almost never the final word on anything significant.

Patterns we use

A few of the patterns the team relies on:

• Specification drafting. Given a ticket and context, AI tools help produce a first draft of a specification — acceptance criteria, edge cases, considerations. A senior reviewer refines and approves before implementation begins.

• Bounded implementation. For well-defined work like refactors and mechanical changes, AI tools handle the bulk of the change end-to-end, with senior review confirming the result against the specification.

• Pre-review automation. Every pull request receives an automated review before a human reviewer engages. This catches consistency issues, missing test coverage, and small inconsistencies — the details easy to miss on a long review.

• Internal tooling. We have built internal tools that streamline onboarding, generate runbooks, and answer routine questions about our codebase. These tools live alongside the work and improve alongside it.

Strengths and limits

The patterns above work well for certain kinds of work and less well for others. We have found AI tooling particularly useful for:

• Detecting drift between specifications and implementations
• Surfacing local consistency issues across a codebase
• Identifying gaps in test coverage
• Executing well-defined mechanical changes

We rely on senior human judgment for:

• Architectural decisions, where the right answer depends on context
• Naming and design choices that require deep domain understanding
• Security determinations, where a flag from a tool indicates a need to investigate but is not itself a conclusion

Continuous improvement

We treat the workflow as a living system rather than a fixed tool. The team shares observations in weekly learning sessions, refines internal tools as patterns emerge, and adapts approaches as the underlying AI capabilities evolve.

The result is a team that delivers more per person, more consistently — without sacrificing the depth of engineering judgment that distinguishes good software work.