The best first automation is rarely the most impressive demo. Look for a recurring task with a clear input, a repeatable result and someone who can check exceptions. Automation may use ordinary rules, AI or both; begin with the process rather than a model.
Choose a bounded, visible task
- It happens often enough to observe, but can be tested on a limited sample.
- The desired result can be described with concrete acceptance criteria.
- A person owns the process and can review or stop the pilot.
- Sensitive decisions and irreversible actions stay outside the first scope.
Draw the current workflow first
Write down the trigger, information used, decisions, hand-offs, output and exceptions. Identify the authoritative system for each field and the people who may access it. This often reveals that a missing rule or duplicated step, rather than a lack of AI, is the main problem.
Use AI only where it adds value
Use deterministic code for stable rules such as checking a date or required field. A model may help classify free text, extract meaning or prepare a draft. Keep permissions narrow and place validation and human review between the model and consequential actions.
Run a small test and decide from evidence
Build a representative test set with normal, incomplete and unusual cases. Compare the pilot with the current process using the same sample: time, corrections, exceptions, review effort, running costs and failure impact. Agree in advance what would justify continuing, changing or stopping. A prototype is not evidence of production readiness.
A useful first workflow has a named owner, authorised data, a manual fallback and a clear way to stop it. Once those pieces are visible, the next step can be chosen with less guesswork.