AI agents for customer support: what actually works in production.

Narrow scope beats broad ambition

The agents that hold up in production are the ones scoped to a specific set of tasks, order status, refund eligibility, appointment rescheduling, rather than an open-ended assistant meant to answer anything. Narrow scope makes the failure modes predictable, and predictable failure modes can be designed around.

Businesses that launch a general-purpose agent on day one usually spend the following months narrowing it after painful edge cases surface in front of real customers. Starting narrow and expanding based on evidence is slower to announce but far cheaper to run.

Escalation paths matter more than the model

Every agent will hit a request it cannot resolve. What separates a support system customers tolerate from one they resent is how gracefully it hands off to a human, with full context transferred, rather than making the customer repeat themselves from zero.

The technical work of passing conversation history, order data and detected intent to the human agent is unglamorous compared to picking a model, but it is what determines whether the escalation feels like a safety net or a dead end.

Grounding in real data prevents confident wrong answers

An agent that answers from a knowledge base connected to live order and account data will say correct things even when it does not fully understand the question. An agent relying only on a general model without that grounding will occasionally invent a return policy that does not exist, stated with total confidence.

This is the single biggest source of support agent failures reported by teams running these systems in production, and it is solved by retrieval against real systems rather than by picking a newer model.

Tone control needs constant tuning

Support agents that sound overly formal frustrate customers who want quick help, while agents tuned to sound too casual can undermine trust on serious issues like billing disputes. The right tone shifts by ticket category, and teams that review transcripts weekly catch tone drift long before it shows up in satisfaction scores.

Measure deflection quality, not just deflection rate

A high percentage of tickets closed without human involvement looks good on a dashboard and can hide a rise in customers who gave up rather than got resolved. Pair deflection rate with a follow-up satisfaction check and a re-contact rate to see whether tickets are actually being solved or just closed.