AI Customer Support Agent

A bot that only answers questions is a search box with a friendlier interface. It can quote a policy, but it cannot look up whether this customer's order actually shipped, or change anything once it has told them the answer. The customer still opens a ticket, and a person still does the lookup the bot was supposedly there to save.

The cost of support is not answering. It is the minutes spent moving between the help desk, the order system and the CRM to find out what is actually true for one specific customer, and then acting on it. A system that skips the acting step has moved the work later, not removed it.

A message arrives, and intent classification decides what kind of request it is before anything tries to answer it — a general question reads and resolves differently from a request tied to one customer's account, and treating them the same is how a bot ends up guessing at account specifics it was never given.

For a general question, knowledge retrieval pulls from the documentation and policies the business already has. For anything account-specific, the agent looks the customer up in the systems that hold the truth: order status, subscription state, ticket history. It answers from what it finds there, not from what sounds plausible.

Where the request calls for doing something — a refund under a set limit, an address correction, resending a receipt — the agent takes that action directly and checks the result before telling the customer it is done. A system that reports success before confirming it has just moved the lie downstream.

What the agent is not permitted to resolve, or cannot resolve with confidence, goes to a person with the conversation, the lookups and the attempted action already attached — not a fresh ticket with none of that context.

Retrieval over the documentation and policy set for the general-question path, and direct connectors into the systems of record — the CRM, the order or subscription platform, the ticketing system — for anything specific to one customer. A mocked or partial connection to those systems produces a bot that is confident and wrong.

A model handles intent and the reply itself; the lookups and the permitted actions run as defined calls against those systems rather than as something the model improvises. Orchestration holds the state of the conversation, retries a lookup that fails, and keeps a record of each turn that a person can review afterward.

Every action the agent can take is listed explicitly, with its own limits — a refund under one amount is a different action from a refund over it, and the agent only ever has the second one if the business decided it should. That list mirrors the authority a human agent already has; it is not broader just because the request came from software.

Every action taken is logged, including the ones the agent considered and declined. Escalation is not what happens when the system fails — it is a boundary the business set on purpose, for the requests it decided a person should always see.

It does not take over decisions that call for judgment rather than a documented rule — a disputed charge, a complaint about safety, anything where the right outcome depends on discretion the business has not written down. Those stay with a person, and the agent's job is to get the context in front of them faster.

It does not work well against documentation that contradicts itself. If two policies give different answers to the same question, the agent will apply one of them silently unless that conflict is resolved first — automating a disagreement does not settle it.