Four phases. A defined end state. No open-ended retainers.

We spend time with the people who actually run the process — not the documentation. We map decision nodes, identify where human judgment is genuinely required, and find the places where a well-specified agent can take over reliably. This phase ends with a written brief that both sides sign off on. If we don't think an agent is the right tool at this stage, we say so.

Before any code: a written specification covering inputs, outputs, tools, model selection, cost constraints, latency requirements, escalation triggers, and acceptable failure modes. Both sides agree on this document in writing. Scope changes after this point go through a formal change request — not a Slack message.

We build against your actual environment. Real data, real APIs, real access patterns. We don't do sandbox demos that break on first contact with production. Integration tests cover normal operation, edge cases defined in the spec, and failure modes. You see test results, not just "it's working".

The agent goes live in your environment. Monitoring is in place from day one. The handover package includes: architecture documentation written for engineers, runbook written for operations, test suite with instructions, and a 30-day stabilisation period where we're available for issues. After that, you have the option of a maintenance plan or full internal ownership.