A support assistant that resolves 61 percent of tickets without a human

Aurient's support team of six was handling 2,800 tickets a month, roughly 70 percent of which were answerable directly from existing documentation. First response time had drifted to 14 hours and the team was losing people to burnout.

An earlier attempt at a chatbot had been withdrawn after it confidently gave customers incorrect billing information.

We built a retrieval-augmented assistant grounded exclusively in Aurient's knowledge base, product documentation and approved policy content. Every answer carries a citation to the source document, and the assistant is configured to decline rather than speculate when retrieval confidence is low.

Billing, refunds and account changes were placed explicitly out of scope: those conversations escalate to a human immediately with the full transcript attached.

The first four weeks went into content rather than code. Aurient's documentation had drifted out of date in several places, and grounding an assistant in stale content would simply have automated the wrong answer. We built a coverage report showing which incoming questions had no supporting document, and their team wrote 40 new articles against it.

Before launch we ran an evaluation set of 300 real historical tickets and measured answer accuracy against what a human agent had actually replied. We iterated on retrieval until accuracy passed the agreed threshold.

The assistant now resolves 61 percent of incoming conversations without human involvement. First response time for the remainder fell from 14 hours to 40 minutes, because the queue is far shorter.

Customer satisfaction rose slightly rather than falling, which we attribute to the citation model: customers can see where an answer came from, and the assistant escalates rather than guessing. Aurient redeployed two support staff to customer success.