Customer Support AI Automation
Agentic support that resolves what it can, escalates what it should, and never pretends to know what it doesn't.
The opportunity
Support tickets had a long tail of repetitive questions buried under genuinely complex ones. Past chatbot attempts solved the wrong half — they answered hard questions badly and easy questions slowly.
How we approached it
- Inventoried ticket types and modeled the cost of getting each one wrong
- Built a tiered agent: fast answers for simple intents, escalation for ambiguous ones
- Hard-coded a "I don't know — handing you to a human" path with full context transfer
- Continuous eval: every escalation is a labeled training signal
What changed
Customers get faster answers on the easy stuff, agents get more context on the hard stuff, and the team stopped tracking deflection rate as the only metric that mattered.
47%
Tickets fully resolved by agent without human touch
–38%
Median handle time on escalated tickets
0
Hallucinated policy answers in 90 days post-launch
Keep reading