This guide is for operators who want to use AI on a real workflow and want to avoid becoming another stalled pilot. Work through the twelve questions below with the people who actually do the work. If you can answer most of them clearly, you are ready to build. If you cannot, you have found your homework.
The questions are grouped using the four functions of the NIST AI Risk Management Framework: Govern, Map, Measure and Manage. NIST released the framework in January 2023 as voluntary guidance, and added a Generative AI Profile in July 2024. You do not need to adopt the whole framework, but its structure is a useful way to make sure nothing important gets skipped.
Map: understand the workflow
1. What exactly is the workflow? Write it down step by step: what triggers it, who touches it, which systems are involved and what "done" looks like. If people disagree on the steps, resolve that first.
2. Where does the time actually go? Time a few real examples. The slow step is often not the one people expect.
3. What happens when it goes wrong today? Understanding today's failure modes tells you what the AI system must handle, and what the cost of a mistake really is.
Govern: decide who owns it
4. Who owns the outcome? Name one person accountable for the workflow's results after launch, not just during the build.
5. Who approves what? Decide which actions the system may take on its own, which need a human to approve, and which it should never take.
6. What data can it use? List the systems and documents involved, who owns them and any privacy, contractual or regulatory limits. Gartner cited poor data quality and inadequate risk controls as two of the top reasons generative AI projects are abandoned after proof of concept.
Measure: define success
7. What is the baseline? Count today's numbers: volume, time per item, error rate, response time or cost. Without a baseline you cannot show improvement.
8. What is the target? Agree on what success looks like and by when. "Unclear business value" is the other reason Gartner gave for abandoned projects.
9. How will you check quality? Decide how outputs will be reviewed, by whom and how often, especially in the first weeks.
Manage: plan for real life
10. What is the fallback? When the system is unsure or unavailable, what happens? The answer should usually be "a person takes over, with full context."
11. How will people learn to use it? Plan training, a feedback channel and a way for users to flag problems in the moment.
12. Who maintains it? Workflows change. Budget time for tuning, updates and new edge cases after launch.
Scoring yourself
- 10 to 12 clear answers: you are ready to build. Start small and integrate with real systems from day one.
- 6 to 9: you are close. Spend a week or two closing the gaps, usually ownership and baseline metrics.
- 5 or fewer: start with workflow mapping before you think about technology.
If you would like a second set of eyes, we run this exercise with teams as the first step of every engagement.
