Resources

Kill criteria in an AI SOW

If “good” and “stop” aren’t written before build, you don’t have a pilot—you have a hope.

Ballast-AI puts kill criteria in paid pilot SOWs by default. Here’s the minimum set operators should insist on—whether they hire us or not.

  • Baseline definition. How the metric is measured, on which population, starting when.
  • Target threshold. What movement counts as success at pilot close—not a vanity dashboard.
  • Go / no-go gates. Explicit checkpoints (e.g., end of week 1 baseline, day 45 adoption, day 90 metric).
  • Control failures. Conditions where work pauses: access denied, audit gaps, unsafe autonomy requests.
  • Scope freeze rules. How change requests are accepted or deferred so the metric stays testable.
  • Handoff definition. What “production” and “maintainable” mean in artifacts and owners.

Related reading: readiness checklist · how we work.

Start with one metric—not a 90-slide roadmap

Tell us where cost, throughput, or labor is stuck. We’ll tell you if a 2–4 week assessment is the right next step—or if it isn’t.