Practical AI · 3-minute read
A bad process used to be limited by how quickly someone could copy and paste. We have made remarkable progress on that limitation.
Give AI a confused instruction and enough access, and the confusion can travel impressively far before lunch. The lesson is not that people are stupid. Smart people inherit messy systems, work around them and eventually forget which workaround became company policy.
Automation can turn that forgotten workaround into a very enthusiastic employee.
First, find the mistake you would be multiplying
Consider Juniper Basin Supply, a fictional oilfield-parts distributor invented for this example. Its office wants AI to draft customer quote emails from a spreadsheet. One column says “available,” but sometimes that means on the shelf and sometimes it means the supplier might have it.
An imagined draft confidently promises Friday delivery. The writing is excellent. The inventory definition has committed a small act of sabotage.
Asking for a more professional tone will not repair that. The team needs to distinguish stock status, supplier confirmation and an approved delivery commitment before an email goes out.
Once those rules are clear, AI could organize the details, flag missing confirmations and prepare a draft for a person to approve. That is a useful job. Inventing the delivery promise is a different job, and nobody hired it for that one.
Give the process a five-minute interrogation
Input: What does each field actually mean? Which source is current?
Rule: What decision should the system make, and what must it leave to a person?
Exception: What happens when information is missing, contradictory or outside the usual case?
Stop: Who checks the result, and who can pause the workflow?
Try a small set of fictional or approved examples before connecting a workflow to live customer communications. Include one easy case, one incomplete case and one disagreement between sources. Count corrections and review time alongside speed.
Try this prompt
Review this proposed workflow: [steps]. Identify ambiguous inputs, unstated decision rules and likely exceptions. For each, explain how an error could spread and suggest a check. Separate tasks AI could draft from decisions requiring a named person. Do not fill gaps with assumptions. Ask for missing information before recommending automation.
The response is a draft process review. Have the people doing the work check it. Keep confidential customer details out of an unapproved tool, and require human approval before quotes or commitments leave the business.
NIST’s AI Risk Management Framework treats context, evaluation and responsibility as connected parts of managing AI risk. My practical version: understand the job before handing it a faster engine.
AI can help expose a bad assumption. That is a much better use of intelligence than distributing it to everyone in your address book.
Your turn: Which recurring mistake in your work would become a bigger problem if it happened ten times faster?
Background: NIST AI RMF 1.0, Core. The company, scenario, checklist and prompt are original editorial examples, not a client case study or a formal NIST checklist.
