Free resource · Version 1.0
A free framework for testing AI ideas before you bet on them.
Most AI pilots don’t stall because the technology fails. They stall because no one defined the problem, agreed what success looks like, or decided ahead of time when to stop. The Canvas takes you through nine sections in four phases—Frame, Design, Evaluate, and Decide—so every experiment ends with a decision, not a shrug.
Download the canvas
Not sure which one? Start with the canvas. If you run out of room, or the team keeps debating a section, move to the workbook.
Save a copy before entering information. For the PDF, use a viewer that supports fillable forms and save your progress as you work.
How to use it
Work through the sections in sequence. Set measures and stop conditions before testing, then make one explicit decision from the full body of evidence.
Ask whether AI is the right answer at all. “No” is a useful result.
State what you expect to happen—and what would prove you wrong—before you run anything.
Track user outcome, AI performance, adoption, and feasibility as distinct signals.
Choose Scale, Iterate, Narrow, Pause, or Stop, then write down why.
Who it’s for
The Canvas works for owner-operators testing their first AI idea and for teams inside larger organizations that need to build a credible case before they scale one.
Free to use and adapt for noncommercial purposes with attribution. Adaptations must use the same Creative Commons license.
Worked example · Fictional
An eight-person online home-goods shop tested AI-drafted replies to routine customer emails for three weeks. A person approved every reply before it was sent.
Routine replies took 26 hours on average and created a growing support backlog.
AI could draft replies, but people retained every decision to approve, edit, or reject them.
Support staff reviewed drafts before customers received anything.
Drafting assistance would reduce reply time from 26 hours toward an eight-hour target.
The team tested shipping, order, and returns questions for three weeks.
Approved policies and FAQs grounded the drafts; unnecessary customer data was excluded.
The team tracked reply time, errors, human edits, and whether staff used the drafts.
Every reply required review, and inaccurate policy guidance triggered correction or a pause.
Reply time fell to nine hours. Shipping and order drafts were strong; returns caused both errors. Decision: narrow, fix the returns FAQ, and retest.
See the method in practice
