Sierra hit $100 million ARR in seven quarters and now exceeds $150 million. Decagon reached $4.5 billion valuation with over 100 enterprise clients. Intercom’s Fin resolves 67% of support queries at $0.99 per resolution. The economics are compelling. But 39% of companies pulled back their AI support bots in Q1 2025 after quality failures. Building a support agent that works is hard. Building one that does not damage your brand is harder.
The Economics Are Undeniable
| Channel | Cost Per Interaction |
|---|---|
| Human agent | $6-8 (loaded) |
| AI agent | $0.50-0.70 |
| Intercom Fin | $0.99 per resolution |
A support organization handling 500,000 interactions/month that automates 60% saves roughly $18-22 million annually. That is not an optimization — it is a structural change.
But the aggregate numbers hide a distribution. The top quartile achieves 70%+ resolution with high CSAT. The bottom quartile creates disasters that go viral. The difference is not the LLM — it is everything around the LLM.
What Separates Success from Failure
Domain-specific knowledge, not generic LLMs. The agent needs codified operational knowledge: exception paths, customer tier policies, known bugs and workarounds, escalation criteria. A generic prompt that says “be helpful” fails on the 30-40% of interactions that require judgment.
Multi-layer guardrails pipeline. Infrastructure-level policy enforcement, PII detection, content grounding against approved sources, escalation triggers. Application-level prompt instructions are insufficient — they can be bypassed by prompt injection.
Latency-aware architecture. Customer support has tight latency expectations. Every middleware layer (guardrails, memory retrieval, tool calls) adds latency. Optimized architectures using Go for the infrastructure layer and streaming responses maintain sub-second perceived latency.
Intelligent escalation. Not just “transfer to human” but handoff with full context: conversation summary, attempted resolutions, customer sentiment, relevant account details. 71% of consumers say they will abandon a brand after a single bad AI interaction — a bad escalation counts.
Measurement from day one. Resolution rate (not containment rate), CSAT per interaction, cost per resolution, escalation quality. Without these, you can’t distinguish success from failure until customers complain publicly.
What This Means for Your Organization
Treat your support agent as a knowledge product — codified domain expertise with embedded quality measurement — not as a chatbot bolted onto an LLM. The organizations getting this right invest as much in guardrails, testing, and measurement as they do in the AI model itself.