A useful assistant needs to recognize when the next step belongs to a person. Drafting a schedule, changing an appointment, and sending a commitment to a customer have different consequences, even when they begin with the same information.
Start with the action
Consider a hypothetical service-business workflow. An assistant organizes job notes, notices a scheduling conflict, and proposes two alternatives. The owner can inspect those options before any customer receives a message. The draft helps the decision move forward without making the commitment on the owner’s behalf.
My design preference is to consider what the action changes, who is affected, what information is missing, and how readily the result can be undone. This is a public design principle, not a description of unpublished controller logic.
Confidence needs context
An answer that sounds certain may still rest on an old record or an unstated assumption. A system should make the source and its date visible when those details affect the decision. If the record is incomplete, a precise question may be more useful than a confident guess.
The NIST AI Risk Management Framework offers a broader, voluntary approach to managing AI risks throughout design, use, and evaluation. I use it as external context for thinking about these choices; this article does not claim a product certification or formal assessment.
Make approval concrete
A helpful approval request should show the proposed change, its destination, and the important consequences. “Proceed?” is hard to evaluate. “Move this visit to Thursday at 10 a.m. and send this message to the customer” gives the person something specific to review.
Approval should also connect to the version being approved. If the proposed recipient or message changes, the earlier decision may no longer cover the action. Keeping a visible record helps people understand what happened.
A practical direction
For the tools I am exploring, I want automation to handle preparation and make decisions easier to inspect. Whether a specific feature achieves that needs testing in its real workflow. An approval button by itself does not establish useful human oversight.
Keep reading
A decision-support assistant · Small businesses and AI
About Tim Hauptrief · Research notebook
Prepared with AI assistance. Featured illustration generated with AI.
Leave a Reply