Start with a piece of work your team already does. It’s much easier to decide whether AI helps when you can watch the work, talk to the person doing it, and compare the results.

That might be sorting incoming support requests, finding an internal document, or drafting a reply to a customer. Pick something you can describe without needing to explain the whole business.

Watch someone do the job

Ask a teammate to walk you through a recent example. Where did they look for information? What did they copy between systems? Which decisions took time, and what did they have to check?

You may find that the first useful change is an integration or better documentation. It’s worth finding that out before committing to an AI build.

Check what you have to work with

You’ll need enough examples to test, permission to use the information, and someone who can tell whether an answer is any good. You’ll also need a person with time to own the workflow once the tool is running.

Think through what happens when the tool gets something wrong. Can someone review the result before it reaches a customer or changes a record? An early project is easier to learn from when mistakes are straightforward to catch and correct.

Keep the first version small

For a knowledge assistant, start with one team and one maintained source of information. Helping support find approved answers, with links to the source, gives you something concrete to test.

Write down who can use it, which sources it can access, and when it should ask for help. You can expand the scope once you’ve seen how it handles the work.

Include the checking time

A draft that takes seconds to generate can still take several minutes to check. Measure how long it takes to get to an answer someone is willing to use, including corrections and review.

Before you start, record how the team handles that work today. Then agree what would persuade you to keep going, change the approach, or stop. Review the results with the people who’ll use the tool - they’re the ones who can tell you where it helps and where it adds work.