Free checklist · runs in your browser
Is your AI workflow ready to ship?
Twelve yes-or-no questions about the mistakes nobody would see. Each yes has a bar to clear. At the end, a scorecard lists every gap, what to fix first, and what you could run in the meantime.
First
How will its work go out?
Readiness depends on how much the workflow does on its own, so the questions that count depend on the design. Answer for the design you plan to ship.
“Goes out” means anything with an effect: an email sent, a record saved, a tag published, a ticket closed.
Your scorecard
How it works
Why it’s built this way
A yes has a bar
Under every question is its bar: one to three things that have to be true, each with its own tick. All ticked is a yes; anything less is a no. A no isn’t a failing grade: it’s the next thing to do, and the scorecard says how to start.
No “partly”
Each bar is either cleared or it isn’t, so “partly” is a no. Survey research finds that an easy way out, such as “not sure,” gets picked to skip thinking the question through. Here, not sure means no evidence, and no evidence means no.
One gap is enough
There’s no “10 of 12, so mostly ready.” Each question guards against a different way of failing silently, so a strong answer on one can’t cover for another. Google’s ML Test Score takes the lowest section score for the same reason.
Ready for what?
A workflow where a person approves everything needs four of these, one of which tests the reviewers themselves. One that lets outputs go out alone needs up to twelve. When the design you planned isn’t ready, the scorecard lists what the cautious one would need, so the work can start while you close the gaps.
Numbers over opinions
Question 5 can only be a yes when your limit and counts are filled in and the arithmetic agrees, and question 11 gets a spot-check plan worked out from your limit. Both use the same likely range as my error-rate calculator and eval harness, so all three agree to the decimal.
Where the questions come from
From failures I found when I measured two AI workflows I’d built, checked against NIST’s AI Risk Management Framework, the EU AI Act, OWASP’s list of LLM risks, and research on checklists, automation and statistics. The method gives the reasoning and every source.
What it doesn’t check
Whether you’re allowed to use the data, security beyond basic prompt injection, legal compliance, or fairness between groups of people. It checks one thing: whether wrong outputs get stopped or caught, and, where some get through, whether you’d know how many.
Private by design
Everything runs in your browser. The link to your answers holds yes and no and your numbers, never the workflow’s name or your notes. Run the tests behind it with npm test.