Automating the Boring Parts Before Reaching for AI
Most "we need AI" requests turn out to be routing, templating and scheduling problems that plain automation solves more reliably.
AI automation is a genuinely useful service line, and it is also the current default answer to problems that do not need it.
Map the loop first
Write down the repetitive task step by step, with who does it and how long it takes. Very often two of the six steps are the entire cost, and both are deterministic — copying a value between systems, or sending the same acknowledgement.
Deterministic work deserves deterministic tools
Routing an enquiry by keyword, generating a monthly report, chasing an unanswered thread: these want rules, queues and templates. They are cheaper to run, easier to debug and they do not surprise you.
Where the model actually earns its place
- Classifying free-text enquiries that resist keyword rules
- Extracting fields from documents in inconsistent formats
- Summarising long threads for a human to act on
- Drafting a first reply that a person then edits
Keep a human in the loop where it costs to be wrong
Draft-and-approve is almost always the right shape for client-facing output. Full automation is for the steps where a mistake is cheap and visible.
Comments
No comments yet — be the first to weigh in.