65% of support tickets now resolved end to end without a human, with satisfaction scores holding.

The situation

The support team was answering the same questions every day. Ticket volume grew with the customer base, the team did not, and the hard tickets — the ones worth a human — sat behind a queue of password resets and billing questions.

They had tried an off-the-shelf chatbot. It answered confidently and wrongly, customers noticed, and it was switched off within a month. That failure mattered: the bar was no longer "deflect tickets", it was "never damage a customer relationship to deflect a ticket."

What we built

An agent trained on their own material — years of resolved tickets, the product documentation, and the tone guidelines their team already wrote to.

Before it answered anyone, it ran against a held-out set of real past tickets, and we compared its answers to what the team had actually replied. That evaluation pipeline stayed in place after launch, so a model change is measured rather than hoped about.

Every answer cites the documentation it came from, so a human agent can check it in one click rather than re-deriving it.

What we deliberately left out

No answering when it is not sure. The agent hands off to a human rather than guessing. Coverage went down; trust stayed intact, which was the actual requirement.

No account actions. It answers questions. It does not issue refunds, cancel plans or change account state — those routes stayed with the team.

No replacing the helpdesk. We integrated into the tool they already used instead of asking a working team to move.

The result

Roughly two thirds of incoming tickets now close without a person touching them, and satisfaction scores did not move — which, given how the previous attempt went, was the number the team was watching. The support team spends its day on the third of tickets that need judgement.