Most stalled AI rollouts follow the same pattern: someone buys licences, sends a welcome email, and waits. A few enthusiasts experiment, everyone else carries on as before, and three months later nobody can say what changed.
Map the work first
Pick one team and list the recurring tasks that eat their week. Good candidates share three traits: they involve reading or writing text, they repeat often, and a person can quickly check the result. Drafting client emails, summarising meeting notes, first-pass document review and turning rough notes into proposals are common examples.
Choose a small number of use cases
Two or three well-chosen use cases beat twenty vague ones. For each, write down what good output looks like, who checks it, and how you will know it saved time.
Then pick the tool
Only now compare tools such as Claude or ChatGPT Team against those use cases. Look at data handling terms, admin controls, and how easily you can share prompts across the team.
Measure honestly
Ask the people doing the work to log time before and during a short pilot. Their judgement of quality matters as much as the minutes saved.