A purchase inquiry goes out to dozens of suppliers, and nobody replies for a while. When should you chase, and whom? This used to run entirely on the operations team's feel. We wanted the system to produce that list every day.
The small task taught two lessons: how to set thresholds from data, and how to encode people's experience as rules the system will not cross.
Lesson one: thresholds are measured, not guessed
How long is "too long without a reply"? We measured every historical gap between sending an inquiry and the supplier's first real reply, excluding auto-responses. The distribution is steep: those who will reply have mostly done so within two hours, and after a day almost nobody replies.
The data confirmed the team's intuition and produced numbers you can write into code: orders due soon enter the list after two hours, distant ones after a day. Why only two hours for near-term orders? Waiting one more hour buys very little additional reply, while for an order due tomorrow an hour earlier is a real opportunity.
Lesson two: whom not to chase
More important than whom to chase is whom not to. Chasing spends our credibility with suppliers, and chasing wrong costs more than not chasing. The team had four unwritten rules; we made them hard ones:
- The order already has enough quotes: don't chase. The goal is quotes, not a complete set of acknowledgments.
- The due date has passed: don't chase. It only shows we weren't watching the order timeline.
- Already waited more than a day: don't chase. The data says nobody answers now; all that's left is annoyance.
- One chase per inquiry round. No response after chasing means stop.
And one that's easy to miss: judge by order, not by supplier. One inquiry often goes to dozens of suppliers and only a few reply. Chasing "whoever hasn't answered" means sending dozens of second emails at once, while history shows most recipients never reply and chasing won't change that. So the unit of judgment is whether this order is short of quotes. If yes, it enters the list; if not, let the silent ones go. That single rule removed the vast majority of unnecessary chasing.
The system's boundary
The system only computes the list. It sends no email and writes nothing to the business system. People still send the chase. Not because the technology can't, but deliberately: chasing spends interpersonal credibility, and that kind of spending should carry a human signature.
Why it's worth writing down
It is a very small feature. But the method reuses across almost every scenario where operational judgment is handed to AI:
- measure the history first so thresholds have a source;
- encode the team's experience as hard rules rather than letting a model "weigh it up";
- pick the right unit of judgment (the order, not the supplier);
- leave credibility-spending actions to people.
Agents take the load; people make the call. Hand over repetitive work one segment at a time, knowing where each segment's boundary is.