Customer Support Agent
Customer support was one of the first places companies pointed LLMs at. Two years in, most of those “AI chatbots” are still rule-based intent classifiers with a language model bolted on top. We build the thing those were supposed to be: a real support agent with tool access, memory, and the good sense to escalate when it shouldn’t answer.
How we build it
Grounded answers, not hallucinations. The agent reads from your actual help center, product docs, and — via MCP — your order / billing / account systems. Every answer carries a citation to where it came from. If the agent can’t find a source, it says so and escalates.
Tool use over text-only. “Where’s my order?” doesn’t get answered by a language model recalling training data. It gets answered by a tool call into your order system, scoped to that customer’s own account. Same for refunds, cancellations, password resets — scoped, logged, and reversible.
Confident escalation. The agent knows what it doesn’t know. Sentiment, policy exceptions, angry customers, legal-adjacent queries — it hands off to a human with the full context attached, not a “please hold while I transfer you.”
Observability and evals. Every conversation is traced. Resolution rate, escalation rate, CSAT, time-to-resolution — dashboards your support lead actually looks at. A regression suite of past tickets catches drift when you upgrade the model.
Transcripts your QA team will read. Structured tool-call logs, not unparseable chat dumps. When something goes wrong, you can tell exactly what the agent saw, what it tried, and why it went the direction it did.