How to Run a Low-Risk AI Workflow Test at Work

Which task is suitable for a low-risk test?

Pick a task with a clear start and finish, enough repetitions to compare, and an output you or a colleague can check before anyone relies on it. A first draft of a summary from approved, non-sensitive notes may qualify. A final customer commitment, hiring recommendation, legal interpretation, or safety decision does not: the cost of an unnoticed error is too high for an informal personal trial. If your workplace has not approved a tool or data type, ask first or use public material in an approved environment.

The International Labour Organization’s 2025 occupational exposure method scores tasks and then examines both their average and their variation within an occupation. Its global analysis therefore describes potential exposure, not whether a particular employer has adopted a tool or whether an individual task can safely run without review. That is why a role label such as “analyst” is a poor experiment brief: an analyst may summarize a public report, reconcile a restricted spreadsheet, and advise a decision-maker, each with different inputs and consequences. ([ILO, “How might generative AI impact different occupations?”](https://www.ilo.org/resource/article/how-might-generative-ai-impact-different-occupations))

Before starting, write one sentence naming the task, permitted input, expected output, and the human check that remains. Set a stop rule too: stop if the tool receives information it should not, invents a material fact you cannot verify, or crosses into work outside the approved scope. NIST’s voluntary Generative AI Profile recommends defining acceptable use, assessing performance in conditions similar to the intended setting, documenting measures, and avoiding general conclusions from narrow, anecdotal checks. It offers risk-management guidance, not proof that this particular two-week design is safe or effective. ([NIST AI 600-1](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf))

Sources: How might generative AI impact different occupations?; Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)

How should you compare the usual workflow with AI assistance?

Use ten working days as a simple schedule: record the usual process on comparable cases during the first few days, try the approved AI step on similar cases in the middle, then repeat and review near the end. This sequence is an editorial recommendation, not a tested duration. If the work varies, alternate similar cases between methods where practical; do not give the tool only the easy examples. Keep the task, acceptance standard, and allowed inputs steady so that a change in results has a plausible connection to the workflow being tested.

For each case, log total hands-on minutes through review and correction, whether it meets the existing quality standard, any factual or policy errors, and the rework needed. A compact rubric might ask whether a summary includes all required points, matches the source, follows the required format, and needs material edits. Record an incomplete result as incomplete rather than treating a polished first draft as finished work. If review time removes the apparent speed gain, that is a useful finding, not a failed measurement.

The strongest counterpoint to a cautious small test is that workplace AI assistance has produced measurable improvements in a real deployment. Brynjolfsson, Li, and Raymond studied a conversational assistant used by customer-support agents at one company and report higher average productivity, with effects differing across workers’ experience and skill. The setting had a defined customer-support workflow and an organizationally deployed tool; it does not establish that an unrelated tool or task will produce the same result, nor does it validate a ten-day personal comparison. It does show why the experiment should measure both throughput and quality rather than assume either that assistance cannot help or that a faster draft means better work. ([“Generative AI at Work,” arXiv manuscript](https://arxiv.org/abs/2304.11771))

Sources: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1); Generative AI at Work

When should you keep, revise, or stop the AI step?

Keep it for this task only if multiple comparable cases pass the same quality and safety bar and show a repeatable net benefit after checking and rework. “Net benefit” may mean less total effort, fewer routine corrections, or a better result on an outcome your work actually values. Choose the measure before the test; otherwise it is easy to move the goalposts toward whichever result looks favorable. If you have too few cases or inconsistent results, the honest conclusion is that the test is inconclusive.

Revise when a specific, fixable issue appears, such as a prompt that omits a required field or a task boundary that includes material the tool cannot verify. Change one thing, note it, and retest comparable work. Stop when errors are consequential, checking costs more than the assistance saves, quality falls below your normal bar, or policy rules out the use. These are local workflow decisions. NIST’s warning against extrapolating from narrow assessments is especially relevant here: a handful of outputs can inform what you do with this workflow, but cannot demonstrate reliable performance across the job.

The verdict is deliberately narrow: a two-week test can help decide whether to retain, revise, or stop one approved AI step. It cannot establish employer-wide adoption, changed labor demand, or a probability that your job will disappear. The ILO separates potential exposure from actual impact, while the workplace study shows that measured effects depend on the particular deployment and worker context. If your task changes or the organization changes the tool, data access, or review rules, revisit the result rather than treating it as permanent. End by asking your manager: “If this specific approved workflow meets our quality bar, can we continue it, and who owns review and accountability?”

Sources: How might generative AI impact different occupations?; Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1); Generative AI at Work

Questions readers ask