B2B SaaS company8 weeks2026Sample

An AI agent that triages support before a human sees it

Inbound support tickets sorted, enriched, and routed automatically — with a hard human gate on anything touching billing or cancellations.

This is an illustrative project archetype showing how we scope and describe work of this kind — not a delivered client engagement. Outcome metrics are withheld until this entry is replaced with a real project.

The challenge

Support volume grew faster than the team. Most tickets were repeat questions already answered in the docs, but every one still consumed a first-response slot. Attempts at a bolt-on chatbot had produced confident wrong answers because it indexed three contradictory versions of the refund policy.

What we did
  • Content governance pass: one authoritative version of each policy, with effective dates as metadata
  • Retrieval filtered on document status before ranking, then re-ranked rather than trusting similarity alone
  • Tool-calling agent wired to the ticketing system, billing lookup, and account status
  • Evaluation harness of 180 real tickets, run in CI on every prompt change
  • Hard human approval gate on refunds, cancellations, and anything customer-facing
Outcome

Repeat questions resolved without a human first-response slot, and the tickets that do reach a person arrive already enriched with account context.

Stack

ClaudeVercel AI SDKpgvectorNext.jsPostgres