Skip to content

AI automation · Starter guide

Where AI automation helps, and where it should stop

Useful automation has a trigger, a narrow task, and a person who handles the exceptions. A chatbot with no boundary is not that.

Added 27 Oct 2026 · 1 min read

Start from the repeated task

Look at the work the team does every day that follows the same steps: tagging an enquiry, copying it into a CRM, answering a question that already has a written answer, or building the same weekly report. Those are candidates. A vague wish to “add AI” is not.

Write the input and the output. If you cannot describe both, the workflow is not ready to automate. The first version should do one of those tasks, not a chain of five.

Keep a person in the sensitive steps

An assistant on the website can answer from pages you approve, and it should hand off when the question leaves that set. Customer-facing messages about price, medical topics, legal topics, or complaints need a defined path to a human.

Do not send documents or personal data into a tool you have not checked. Use API settings that avoid training on your content where the provider offers that choice, and write down what was decided.

Make it possible to pause

Every workflow needs an owner and an off switch. If the CRM field names change or the model returns nonsense, someone should know how to stop the automation without calling an emergency.

Judge the work by the hours it returns and the errors it introduces. A report that looks clever but is wrong every Friday is not a saving.

This starter guide was written for the Nerix insights library. It should be updated when the advice changes. It is not a client story and it does not report results.