Designing service triage that scales without frustrating customers

Service automation fails in a specific way: it handles the easy cases well, mishandles a minority, and the mishandled minority generates the complaints that define the programme. The design question is not how much to automate but where the boundary sits and how quickly it can be moved.

Start from the volume, not from the technology

Twelve months of case history usually shows that a small number of intents account for most volume. Classify that history first. The intents worth automating are high-volume, low-variation and low-consequence-if-wrong. Everything else stays with people until the evidence changes.

Three boundary rules

1. Confidence thresholds route uncertain cases to human triage rather than guessing. An unclassified case is a manageable cost; a confidently misclassified one is not.
2. Escalation is always available and never hidden. Customers who cannot reach a person escalate through complaint channels instead, which is more expensive for everyone.
3. Sensitive circumstances such as vulnerability indicators, complaints and regulatory matters bypass automation entirely by rule, not by model judgement.

Measure the things that reveal failure

Agent assistance before agent replacement

Drafting support for human agents is usually the higher-return starting point. It affects every case rather than a subset, keeps a person accountable for the response, and produces the review data needed to decide which intents can later be handled end to end.