Map the work before adding AI
Write down the trigger, inputs, decisions, systems, output, owner and exception path. If nobody can explain the current process consistently, automating it may scale confusion.
Look for repetition with judgment
Useful candidates include classifying enquiries, drafting structured responses, extracting information from standardized documents, summarizing updates, preparing first-pass reports and routing tasks.
Keep deterministic work deterministic
Not every workflow needs a language model. Fixed calculations, validations and simple rules are often better handled with conventional software.
Design the exception path first
Ask what happens when confidence is low, data is missing or a downstream system fails. A production workflow needs escalation, logging and human override.
Measure before and after
Baseline time per task, volume, error rate and turnaround. Compare the same metrics after automation plus the amount of human review required.
Automate the repetitive layer so people can spend more attention on decisions, relationships and exceptions.