AI support triage system for a 200-seat SaaS company
A tier-1 triage system that classifies, drafts, and routes support tickets — so humans approve instead of typing.
Tier-1 was drowning
A 200-seat SaaS ran a tier-1 support team that triaged every ticket manually. Average response was 4 hours, weekends collapsed to 12+, and quality varied by agent.
- 80% of tickets fell into 5 known categories
- Refunds, access issues, and bug reports dominated the queue
- Macros were stale, so agents wrote replies from scratch
The cost of leaving it alone
Slow first response was the single biggest driver of churn-risk conversations. Every hour of delay generated follow-up tickets, and weekend backlogs poisoned Monday capacity.
Replace the bottleneck, not the team
A router that classifies, drafts, and routes. Humans approve or edit; the model never sends without sign-off on categories where it scores below threshold.
- Claude handles classification and reply drafting, citing the KB article used
- n8n orchestrates: webhook → classify → draft → route
- Postgres stores ground truth for nightly eval runs
- Slack DM to on-call when confidence drops below 0.7
Stack: Claude · n8n · Postgres · Slack
How it was built
- Week 1–2: audit of 90 days of tickets; category taxonomy and eval set built from real data
- Week 3–4: thin slice live on one category with human review on 100% of drafts
- Week 5–6: remaining categories enabled as eval scores cleared the bar
- Week 7–8: hardening — confidence routing, weekend coverage, dashboards, team training
What the numbers say
What happened next
Nine months in, the system still runs on a light retainer: nightly evals catch drift, and new product features get folded into the taxonomy in a monthly review. The support team shrank by attrition, not layoffs — and the model cites its sources, so agents catch its mistakes faster than they caught each other’s.
This system is an example of AI Agents & Internal Assistants work.
Need a similar system?
Let's talk through your version of this — same architecture thinking, scoped to your operations and tools.
30 minutes · no pitch deck · reply within 24h if you write instead