Score the workflow before you build the agent.
Use 16 plain-language checks to choose one worthwhile AI workflow, expose the missing decisions, and select the next step that fits the evidence.
How to score
0 means no evidence yet.
1 means partly true or unclear.
2 means clear and evidenced.
The worksheet
Make the next step visible.
Rate each line against one named workflow. Your score is a prompt for action, not a verdict.
01
Opportunity
One workflow is named.
Point to the current process, not a broad AI ambition.
The problem is observable.
Show delay, rework, cost, risk, or missed capacity.
A decision owner exists.
One person can approve the use case and resolve exceptions.
The result is measurable.
Choose a before-and-after measure and review date.
02
Workflow
The current steps are mapped.
Make inputs, handoffs, exceptions, and output users visible.
Data and access are available.
Name what the tool needs and who can grant permission.
Human review is placed.
A named person checks material output before it changes the work.
The fallback is clear.
The team can continue safely when the tool is wrong or unavailable.
03
Governance
Data boundaries are written.
Allowed and prohibited inputs are clear to users.
Approval and permissions are recorded.
Purpose, owner, system access, and limits have a short record.
Material output leaves evidence.
The team can show the input, review, change, and final decision.
Stop and incident paths exist.
Someone can halt use, report a problem, and preserve the record.
04
Delivery
A sponsor can fund the next step.
Budget authority and the procurement route are known.
A workflow owner has time.
The responsible person can test, answer questions, and accept the work.
Users can learn the new process.
Training, support, and exception handling have an owner.
A decision date is set.
The pilot ends with a stop, revise, or continue decision.
Use boundary
This scorecard is a planning aid. It is not an audit, certification, legal opinion, security review, benchmark, or guarantee. The result bands are decision prompts. They are not derived from customer outcomes.
The method is consistent with the context-first logic in the voluntary NIST AI Risk Management Framework, including intended purpose, users, impacts, limits, and the decision to continue or stop. It does not claim NIST compliance.