How to choose your first business AI automation

Start by watching a task happen. Ask the person doing it to show you the inputs, the decisions, and the finished result. A process that sounds simple in a meeting may involve exceptions that only become visible when you work through an example.

Pick a task you can describe

Look for work that repeats, has a clear finish, and can be reviewed before it affects a customer or another system. Drafting an internal weekly update could be a starting point. Automatically approving payments would require a much more demanding design.

For each candidate, ask: how often does it happen, how long does it take, and what would a mistake cost? A task that takes five minutes occasionally may not justify building and maintaining an integration.

Map the smallest useful workflow

The first pilot could use a manually supplied file. Connecting live systems can come later if the drafting step proves useful. If the task is simply adding numbers or moving a fixed field, a formula or conventional automation may be enough.

Agree on what improvement means

Before the pilot, record the time spent preparing and checking several updates. During the pilot, record the same things, including corrections and troubleshooting. Compare the total effort, rather than only how quickly a draft appears.

Keep a short list of errors: missed items, invented details, stale information, or formatting problems. Decide which errors are acceptable to correct and which should stop the workflow.

Give someone ownership

Name the person who reviews the output, maintains the instructions, and pauses the workflow when it fails. For a connected system, agree on access permissions and how to return to the original process.

This focus on responsibilities and evaluation is consistent with the broader risk-management approach in NIST’s AI Risk Management Framework. The example here is a suggested starting exercise, not a client case study or a promised result.

At the end, make a clear decision: keep it, revise it, or stop. A useful pilot gives you evidence for that decision before you expand the work.