Map the task first
Choose one repeated task and observe how it is done. Note the inputs, decisions, exceptions and final output. How often does it happen? Who checks the result? Where does a mistake become expensive?
A workflow that moves a completed form into a CRM may only need an API connection and reliable validation. A workflow that extracts information from varied documents may benefit from a model, followed by checks and human review.
Keep predictable rules predictable
If a rule can be stated precisely and the input is structured, start with ordinary software. It is easier to test the rule “send a reminder three days before the due date” than to ask a model to decide whether a reminder feels appropriate.
AI is more useful when interpretation, summarisation or drafting is part of the task. Even then, show the source, distinguish uncertain output and define who makes the final decision. Do not treat confident wording as evidence of accuracy.
Pilot before expanding
Test representative examples, including awkward cases, against a clear definition of acceptable output. Compare the review effort with the current process. Include model usage, integration maintenance and exception handling in the cost.
Agree which information may leave your systems, review provider retention settings and restrict access. A sensible pilot has a fallback, a usage limit and a stop condition if it cannot meet the required standard.