Start with a process, not with the technology
You don't need "an AI strategy" to get started: you need one specific process that's repetitive, slow or dependent on a single person today. A customer service agent, a sales follow-up agent or a reconciliation agent are real starting points; "using AI in general" is not.
Check whether the process is already clear
An agent can only follow rules that already exist. If two people on your team handle the same case differently, that disagreement needs to be resolved beforehand — not left for the agent to decide on its own which approach is correct.
Check which systems you already use
A useful agent connects to your actual CRM, calendar or ecommerce platform — it never works from a separate copy of your data. If you don't currently know for certain which tool is the source of truth for a piece of data, resolve that before automating anything around it.
Define who approves what
Before granting real access, decide what an agent can do unsupervised and what always needs a person to confirm it. Booking an appointment can be fully automated; a cancellation with a refund probably shouldn't be, at least at first.
Signs that now isn't the moment yet
If no one on your team can explain the process in five steps, if the data lives in loose spreadsheets with no clear owner, or if the real priority is urgent and can't support a trial period, it's worth fixing that first. An agent built on an unstable foundation fails more visibly, not less.