Readiness is a property of the workflow, not the model. The model will keep improving. Your workflow won't, until someone fixes it.
When an AI initiative stalls, teams blame the model: not accurate enough, hallucinated, didn't understand our business. Then a better model arrives and the initiative stalls again. The workflow it was dropped into was never ready to receive it.
"Is the AI good enough?"
Leads to benchmark debates and waiting for the next release. Nothing about your operation changes.
"Is the workflow ready?"
Leads to six testable properties of the work itself. Each one is fixable without waiting for anyone's roadmap.
A workflow is AI-ready when it passes all six. Each test takes minutes to run and points directly at what to fix.
Can anyone describe this workflow end-to-end?
Fails when: the process lives in one person's head, or every case is "it depends."
Do the inputs and outputs exist in systems?
Fails when: the real workflow actually runs over calls and tribal knowledge.
Can a human check the output faster than producing it?
Fails when: verifying the AI's answer takes as long as doing the work.
When the AI is wrong, can you catch and undo it?
Fails when: a single bad output is irreversible or compliance-critical with no review step.
Does one accountable person have authority to change this workflow?
Fails when: the workflow crosses three teams and belongs to none of them.
Does this work happen often enough to matter?
Fails when: it's a quarterly task someone does in an afternoon.
In practice, most workflows fail the same two tests: Verifiable and Owned. Teams obsess over data quality. What actually kills AI initiatives after launch is output nobody can check and a process nobody owns.
Score the six tests and every workflow lands in one of three bands — each with a different correct move.
Not Ready
Fails Defined, Digitized, or Owned
The workflow itself is broken or invisible. Bolt AI onto it and you automate the chaos: faster wrong answers, at scale.
Move: fix the process first. This is a process project, not an AI project.
Ready With Work
Passes structure, fails Verifiable or Recoverable
The workflow is sound but unsafe to augment. Errors would slip through, or couldn't be undone.
Move: design the review checkpoint and rollback path, then proceed. This usually takes weeks.
Ready
Passes all six
Defined, digitized, checkable, recoverable, owned, and frequent. AI lands on solid ground here.
Move: proceed, with the baseline written down first (see Framework 01).
Amber is not a rejection, it's a short to-do list. Most of what gets called an AI failure was an amber workflow treated as green.
One workflow at a time. A department is too big to test.
With the people who do the work. Managers describe the process as designed, not as it runs.
Red, amber, or green. A "probably" on any test is a fail.
Each failed test names its own remediation. Re-test, then build.