Shadow Run · Field sheet

Before your AI
touches customers.

The one-page cheat sheet for running an AI candidate beside your live workflow.

Six checks. Five stop conditions. A place to name who fixes what.

Free to download. Print it, mark it up, and use it in your next review.

The review, on one page.

Score the case: Answer quality, source grounding, escalation judgment, tone and format, permissions, and handoff quality.
Know when to stop: Five conditions to hold the affected scope, with room to record the evidence.
Give the failure an owner: Knowledge, agent setup, product promises, routing, policy, and tools each need someone to fix what failed.
Make the call: Record ready, hold, or constrained scope, with an owner and a date. Keep the counts behind the percentages.

Start on Monday

Use it before a first rollout, and again when a model, tool, policy, or knowledge source materially changes.

The test shows how the candidate behaved on those cases. Keep watching customer outcomes after release.

Pick one workflow and name its owner.
Write the acceptance bar, limits, and stop conditions before testing.
Mix routine and risky cases. Keep fresh cases aside, and use only the context available when each request arrived.
Disable customer messages and live account changes through tools and permissions.
Review the results against evidence. Score each case and record what failed.
Give each failure an owner. Fix and retest, including load, routing, shutoff, and fallback.
Record the decision and allowed scope. Start a small rollout of the work that passes, and keep reviewing.

The sheet is ready to use on its own. The Second Edition develops the method further.