Key takeaways
- Define the task, user, decision, inputs, output, authority, and unacceptable outcome before comparing models.
- Score the workflow against a measured baseline and a viable non-AI alternative—not against a staged demo.
- Prefer narrow, frequent, reversible assistance when evidence is thin or failure is expensive.
- Use assist, approve, automate, defer, and refuse as explicit operating choices.
01
The backlog is usually a list of technologies pretending to be a list of jobs
“Add an agent,” “personalize with AI,” and “automate content” sound like projects, but they do not say who is trying to do what. A useful brief names the person, trigger, input, decision, output, destination, frequency, current method, and consequence of a wrong result. Without that boundary, a successful demo can quietly become permission to redesign half the revenue system.
NIST’s AI Risk Management Framework asks organizations to define the business context, task, assumptions, limitations, risk tolerance, and viable non-AI alternatives. That is not paperwork for its own sake. It is how a team separates a small drafting assistant from an autonomous workflow that changes spend, contacts a customer, or writes to the CRM.
Swipe to compare every column
| Question | Useful evidence | Warning sign |
|---|---|---|
| What job changes? | Observed workflow and baseline | A feature looking for a problem |
| Who bears a bad result? | Named user and affected customer | “The business” in the abstract |
| Can it be reversed? | Rollback, correction, and audit path | Irreversible external action |
| What beats AI today? | Human, rule, template, or process comparison | Demo compared with nothing |
02
Rank value and exposure separately
Estimate volume, time recovered, delay removed, quality improvement, revenue relevance, and learning value. Then score privacy, legal exposure, brand harm, customer impact, integration reach, action authority, detectability, and recovery cost. Do not subtract one total from the other; a large upside does not make an unacceptable failure acceptable.
Confidence deserves its own column. A frequent manual step with clean inputs and a measurable result is easier to test than a rare “strategic” decision whose quality becomes visible months later. When confidence is low, buy evidence with a small pilot instead of buying scale.
03
Choose an operating mode, not a binary yes or no
Use five dispositions: assist a person, prepare an action for approval, automate inside a narrow boundary, defer until prerequisites exist, or refuse because the use is inappropriate. The same model can be acceptable for summarizing internal notes and unacceptable for deciding which vulnerable customer should receive a high-pressure offer.
Start new logic in observe mode when possible. Record what it would have recommended, compare that with the team’s decision and later outcome, and inspect disagreement. Expand authority only when the evidence supports the next step and the recovery path has been exercised.
04
Revisit the decision after the novelty wears off
Assign an owner, evaluation set, review date, cost ceiling, quality threshold, incident trigger, and retirement condition. NIST’s framework treats monitoring, change management, override, recovery, and decommissioning as part of deployment—not as cleanup after a problem.
A sensible portfolio will contain fewer dramatic launches and more bounded improvements: cleaner intake, faster research, better routing suggestions, consistent source checks, and visible exception queues. That may look less futuristic in a keynote. It is more likely to survive a quarter.
Primary sources and further reading
Use the source material to validate details against your own context and current platform configuration.
This field note follows the XenGrowth editorial policy: primary sources where available, visible limitations, material review dates, and no invented first-hand experience.
Stay with the problem
