It does not start with "let's replace the process". It starts with a single work stream, measured honestly — and with a go/no-go decision based on numbers, not enthusiasm. Three phases, each with a hard exit condition.
Every transition between phases is a decision by the organization's owner — made on evidence from the previous phase, just like the gates inside the process.
A review of what is there — without changing anything.
One team, one work stream, production tasks — not a sandbox.
Expansion where the pilot proved value — in order of risk.
AI SDLC does not remove people from the process — it moves them where they are irreplaceable: goals, decisions, risk assessment, acceptance of results. Honestly: daily work changes significantly.
Less hands-on implementation, more task definition, review of agent work, and technical decisions the process will not make on its own.
Agents run the tests; a human designs what "verified" even means in a given domain — criteria, edge cases, checklists.
Instead of status updates and task assignment: defining goals, decisions at checkpoints, and accepting results — with a full audit trail.
The operator role is the heart of the adoption. The operator expresses needs in the language of goals, makes the decisions the process does not take away from them, and accepts results on evidence. They do not need to read code — they need to understand what they want and be able to say "no". Preparing operators is the largest part of the people work — which is why we start it in the assessment phase, not in the rollout.
Stages, gates, and contracts tailored to your workflow and requirements (including audit ones) — not a copy of someone else's process.
Integration with your issue tracker, repository, and CI; configuration of agents, gates, and the evidence trail on your infrastructure.
Preparation of operators and the team, jointly running the pilot, then oversight decreasing until self-sufficiency.