Cutting customer-support resolution time 60% with a triage agent

The challenge

A Series-C SaaS company's customer-support team was drowning. Ticket volume had tripled in twelve months while headcount stayed flat. Average first response was 2.3 days. The pattern was painful but obvious: 70% of tickets were resolvable by pointing customers at existing documentation — but finding the right doc took support reps an average of 11 minutes per ticket.

The client had already tried two off-the-shelf chatbots. Both shipped well, demoed beautifully, and quietly died in production within six weeks: hallucinated answers, no escalation policy, no integration with their actual ticketing stack.

What we built

A production-grade triage agent embedded directly in their support workflow:

• Ingestion pipeline for their docs, runbooks, and historical tickets, with hybrid retrieval (vector + keyword + product/version filters).
• Triage agent that drafts a candidate response with cited sources, classifies the ticket by complexity, and routes 1-of-3 ways: auto-resolve, human-in-the-loop review, or direct escalation.
• Integration layer wired to Zendesk, their internal product API, and a Postgres-backed audit log for compliance.
• Eval harness running on every prompt change and model upgrade — ~400 golden tickets, scored on accuracy, citation quality, and escalation appropriateness.

We deliberately did NOT build: a customer-facing chat interface (their support team still owned the relationship), an agent that wrote code (out of scope), or anything that touched billing decisions (regulatory complexity not worth the risk).

Outcomes (after 90 days in production)

• 60% reduction in average ticket resolution time
• $1.4M annualized support cost reduction (avoided four scheduled hires)
• 94% of agent-drafted responses approved by the human reviewer with no edits
• Zero customer escalations traced back to a hallucinated source — citation enforcement was non-negotiable

How we approached it

Two-week discovery. We sat in on three days of support standups and shadowed reps on real tickets before writing a line of code. We came back with a written recommendation that the right first build was internal-facing triage, not customer-facing chat — the higher-leverage win with lower risk.

Six-week build phase. Every Friday: demo to the client team plus a stakeholder from support. Every Monday: course-correct based on what we'd learned the previous week.

Two-week production rollout. Started with 10% of incoming tickets, scaled to 50% after week one, 100% by end of week two. Eval harness ran on every change.

"What we appreciated most was the willingness to tell us what NOT to build. Every other vendor pitched maximum scope. Intellineo scoped down twice during discovery." — VP of Customer Success, client team

What we learned

Three patterns that have since shown up in every agentic project we've shipped:

1. The eval suite is the deliverable. The agent itself is replaceable — every model upgrade is a chance to swap it. The eval suite is what makes that safe.
2. Human-in-the-loop is not a fallback, it's the default. Auto-resolve is for the 30% of clearest cases. The other 70% should always involve a human — but with the agent doing 90% of the cognitive work upfront.
3. Citations are non-negotiable. Every response the agent generated had to trace back to a source document. Hallucinated sources would have killed the project in week one.