Research Note
First Workplace AI Task Selection Record
A first task is suitable when it is permitted, bounded, reversible, observable, repeatable, measurable, reviewable, and low enough in consequence for the planned controls
In this article
First Workplace AI Task Selection Record
Decision principle
A first task is suitable when it is permitted, bounded, reversible, observable, repeatable, measurable, reviewable, and low enough in consequence for the planned controls. "Low risk" is not a property of summarization, drafting, or classification in the abstract. It depends on the data, audience, decision, workflow, and failure.
NIST's AI RMF Core calls for a defined context, intended purpose, risk tolerance, task, human oversight, benefits, costs, third-party components, and impacts. The record supports task selection without certifying a product or use.
Removal gates
A candidate leaves the first-task list when the worker lacks permission, the input contains unapproved data, the output directly decides a person's rights or employment, an error is hard to detect, the result is difficult to reverse, no qualified reviewer exists, or the organization cannot stop and return to the prior process.
Tasks involving legal interpretation, diagnosis, financial approval, safety control, personnel evaluation, customer commitments, production code, credentials, regulated records, or public claims generally need more than a first experiment.
Positive selection factors
The strongest candidate has a clear start and finish, known source material, a stable acceptance standard, enough ordinary cases for comparison, and a reviewer who already understands the work. A draft created from approved material can fit. An open-ended request to "run the department" cannot.
Frequency is useful only when repeated use creates enough evidence and the consequence remains bounded. High-volume error can create high-volume harm.
Required record
The selection record should name the task, owner, current process, input source, allowed data, output destination, reviewer, acceptance standard, ordinary failure, harmful failure, baseline measure, test period, stop rule, and decision date.
The task should be written narrowly enough that adjacent uses remain prohibited until separately reviewed.
Sources
Follow the evidence.
- NIST AI RMF Measure guidanceairc.nist.gov
- ftc.gov: ai companies uphold your privacy confidentiality commitmentsftc.gov
- youtu.be: 0cC1Ez33ryIyoutu.be
- daltonanderson.ghost.io: ai in the workplace a practical guide to get starteddaltonanderson.ghost.io
- NIST AI Risk Management Frameworknist.gov
- NIST AI Resource Centerairc.nist.gov
- eeoc.gov: prohibited employment policiespracticeseeoc.gov
- eeoc.gov: us eeoc and us department justice warn against disability discriminationeeoc.gov
- nber.org: w31161nber.org
- open.spotify.com: 7LIXDoSM2gG97vFGftskQsopen.spotify.com
- NIST Privacy Frameworknist.gov
- nber.org: w33795nber.org
- eeoc.gov: strategic enforcement plan fiscal years 2024 2028eeoc.gov
- NIST Generative AI Profilenvlpubs.nist.gov
- ftc.gov: start security guide businessftc.gov
- dol.gov: ten 07 25dol.gov
- hbs.edu: dell acqua et al 2026 navigating the jagged technological frontier 5c589c8c fbb5 458f b285 c944746cd717hbs.edu
- cisa.gov: cisa and uk ncsc unveil joint guidelines secure ai system developmentcisa.gov