An AI agent that closes support tickets.
A B2B SaaS team was drowning in repetitive tickets. We built a retrieval-grounded agent that drafts replies, pulls the right docs, and resolves routine requests end to end — escalating only what genuinely needs a human.
A support queue growing faster than the team
The company was scaling fast, and support volume scaled with it. Roughly 70% of incoming tickets were variations of the same few dozen questions — billing, onboarding, integration setup — yet every one waited in the same queue behind genuinely hard problems.
Agents spent their day copy-pasting from a sprawling help center, response times crept past a day, and the best people burned out on work that didn't need them. Off-the-shelf chatbots had been tried and quietly switched off: they hallucinated, couldn't see customer context, and made things worse.
Grounded answers, not confident guesses
We started narrow on purpose. Instead of "an AI that answers everything," we scoped a single, measurable goal: safely resolve the top 30 ticket types with answers grounded in the company's own documentation and account data.
Every response is built from retrieved, cited sources — never the model's memory. If the agent can't ground an answer with high confidence, it doesn't guess: it hands off to a human with a full summary and its best draft attached, so the agent's work is never wasted.
The system
- A retrieval pipeline over the help center, changelog, and internal runbooks, re-indexed automatically on every docs update.
- An agent loop with tool access to the billing and account APIs, so answers reflect the customer's real state.
- A confidence gate and human-in-the-loop handoff that escalates with a drafted reply and context summary.
- An eval suite of 400+ real, labelled tickets that runs on every change — no prompt ships without passing.
- A live dashboard tracking resolution rate, deflection, cost per ticket, and escalation reasons.
"It's the first automation our support team actually trusts. It does the boring 60% flawlessly and hands us the hard 40% already half-solved."
Faster answers, happier teams
Within six weeks of launch the agent was resolving 61% of incoming tickets without a human, and average resolution time across the whole queue dropped 68% as agents were freed to focus on complex work.
Crucially, satisfaction went up, not down — customers got accurate, instant answers at 2am, and the human team got their hardest tickets pre-researched.