01
Scoping
We start by defining what “wrong” looks like, because you cannot guard against a failure you haven’t named. For an automation that means the malformed record, the rate limit, the duplicate event. For an agent it means the question nobody anticipated and the answer that is confidently incorrect.
Scope is written as what is in and what is explicitly out. The out-of-scope list is the more useful half — it is what prevents a fixed price from becoming a fight later.
02
Evaluation before implementation
On AI work, the evaluation suite is built before the agent. A graded test harness, run on every change, so quality is measured rather than felt.
You set the bar with us. We ship when the numbers clear it — not when the demo looks good.
03
Guardrails
Input validation, output constraints, refusal behavior, and cost and token ceilings. Human review on the decisions where being wrong is expensive.
We do not build fully autonomous systems for decisions with real financial, legal or safety consequences. That is a standing refusal, not a negotiating position.
04
Failure handling
Every external call gets error handling and retries. Everything that can fail silently gets a dead-letter path. Anything running unattended gets monitoring on its failure modes, not just its uptime.
This is roughly 40% of the work on an automation and it is the part most vendors skip, because it is invisible in a demo.
05
Named accountability
One senior engineer owns the engagement. They sign off on architecture, evaluation design, the security surface, the handover documentation, and the decision to go live.
You know who that is before you sign. If they change, that is a conversation with you — not an internal reassignment.
06
Handover
Architecture, decisions, prompts and eval cases where relevant, a credentials map, and a runbook. Written to be read by someone who wasn’t there.
Handover documentation is a deliverable, not a favour. Taking the work in-house later is a legitimate outcome and we build for it.
07
After launch
Thirty days of post-launch fixes are included on every engagement. Beyond that, a care plan or you run it — both are fine, and we will hand over the monitoring configuration either way.
On a plan, we contact you first. If something breaks, you hear it from us with the cause, the fix, and what we changed so it doesn’t recur.