Software that answers and acts on its own — within the limits you set for it.
AI agent. An AI agent is software that uses a large language model, connected tools and a knowledge source to complete tasks with little or no human help. In customer support, an AI agent reads an incoming question, retrieves relevant answers from a knowledge base, and either resolves the request or routes the conversation to a person.
Three layers make an AI agent, and only one of them is the model. The language model reads and writes; retrieval grounds what the model says in your actual documentation, so answers come from your policies rather than the model’s general training; and tools let the agent do things — look up an order, update a subscription, tag and route a conversation. A defined escalation path completes the design: the conditions under which the agent stops and hands the conversation to a human, with its context attached.
The under-appreciated consequence: an AI agent’s ceiling is set by what it can retrieve, not by the model. Teams shopping for a smarter model while their help center is thin are optimizing the wrong layer — retrieval from missing or stale articles produces fluent, confident, wrong answers, and the fluency is what makes them dangerous. The fastest way to improve an AI agent is usually to improve the documentation it reads.
Done well, an AI agent absorbs the repetitive majority of support volume at any hour, with no queue, and leaves humans the conversations that need judgment — which is a better job. Done carelessly, it produces wrong answers at scale, each delivered with perfect confidence, and every failure is a customer-facing event.
One terminology trap is specific to support: “agent” has meant a human being in this industry for decades, so “AI agent” invites genuine confusion in metrics, staffing conversations and customer expectations. Be precise internally — and honest externally. An AI agent that lets customers believe they are talking to a person loses their trust the first time the illusion slips, and it always slips.
Start where the documentation is strongest, not where the volume is highest — the agent can only be as good as what it retrieves, so its first assignments should be the intents your help center already answers well. Define escalation triggers before launch: low confidence, negative sentiment, an explicit “let me talk to a human,” and any intent on your never-automate list.
Then supervise like it is a new team member, because it is. Read a sample of its transcripts every week; measure not just how many conversations it resolves without human touch, but how many of those “resolved” conversations come back within a few days — the comeback rate is the honest version of the resolution rate.
An AI agent is the deployed form of several ideas working together — these entries cover the autonomy behind it, the oversight around it, and the content underneath it.
Give your customers faster answers and your team their evenings back.
Desktop and mobile. Free for solo.