How We Built a Support AI Agent in 8 Weeks

Northline SaaS came to us after a pricing change doubled support volume. Macros were stale, contractors could not access internal runbooks safely, and leadership wanted deflection without damaging CSAT.

We scoped a single outcome: resolve tier-1 billing and how-to tickets with cited answers, escalate everything else with full context. No billing changes without human approval—non-negotiable from week one.

Weeks one and two were knowledge audit, not coding. We indexed 340 help articles, removed contradictory macros, and built an evaluation set from real transcripts. Policy matrices defined PII handling, refund boundaries, and when confidence scores must route to people.

Weeks three through five shipped the agent, RAG pipeline, and Zendesk bi-directional sync. Every reply included source citations so agents and customers could verify answers. A QA harness replayed the eval set on every deploy.

Weeks six through eight were shadow mode, then 25% → 100% traffic with weekly reviews on live conversations. We tuned prompts from won escalations—not from synthetic demos.

Results at launch: 58% auto-resolution, 72% faster first reply, CSAT held at 4.6★. The project finished in eight weeks because discovery locked scope early and integrations were prioritized over UI polish.

If you are planning a similar agent, copy the sequence: audit content, define permissions, integrate where work already happens, eval before scale, and measure deflection plus satisfaction—not chat volume alone.