The front door is the worst place to start

Customer-facing deflection is the most visible AI support project and the hardest one to get right, because every failure is witnessed by the person you were trying to help. Teams start there because it is legible to executives, then spend months defending it.

The three positions below are less visible and more forgiving. Each one produces measurable savings without putting a model between an angry customer and a resolution on day one.

1. Triage, before a human opens the ticket

Classifying, tagging, prioritising, and routing incoming tickets is a judgment task that a model does well and a queue does badly. It is invisible to the customer, the failure mode is a mis-route that a human immediately corrects, and every correction is a labelled example.

Routing accuracy is also easy to measure against what your team already does, which makes this the cleanest first project in the whole category.

2. Drafting, in front of an agent

A drafted reply that an agent edits before sending captures most of the speed benefit with none of the deflection risk. The human stays accountable for what goes out, and the edit distance between draft and sent message is a precise, continuously generated quality metric.

Draft-and-edit gives you the speed of automation and the accountability of a person. Most teams skip straight past it.

3. Knowledge, behind everything

Support teams lose more time hunting for the right answer than typing it. A retrieval layer over your documentation, past tickets, and release notes speeds up every agent on every ticket, and it improves the two positions above at the same time by giving them better context.

It also surfaces documentation gaps as a side effect: the questions retrieval keeps failing on are precisely the articles nobody has written.

Then, and only then, the front door

By the time those three are working, you have a measured routing model, a corpus of approved replies, and a retrieval layer that has been corrected by humans for months. Customer-facing deflection built on that foundation is a different proposition from one built on a vendor demo.

The sequencing costs you a quarter and saves you the public failure that sets these programmes back a year.