The 30-second answer
An AI agent is software that handles a customer message end to end. When the customer types "where's my order?", the agent doesn't just reply "I'd be happy to help" — it pulls the order from your back end, checks the carrier, and answers with the real status. Same chat window as a chatbot, fundamentally different capability.
What makes it an "agent" rather than a chatbot
Three things, together:
- It uses a large language model — the same kind of AI that powers ChatGPT. Not pattern-matched scripts.
- It can take actions in your systems — issue a refund, change a shipping address, escalate to a person — not just send text.
- It can chain steps together — look up the order, check the return policy, then issue the refund — in one flow, without someone scripting each path in advance.
Strip away the marketing and that is the definition every credible vendor lands on.
The vendor landscape
Companies positioning explicitly as AI-agent vendors today: Intercom Fin, Decagon, Sierra, Ada, Forethought, Lorikeet, Ultimate. Each handles a specific shape of conversation well; none is best at everything.
The platform players — Zendesk AI, Salesforce Agentforce, Freshworks Freddy — ship their own AI agents inside the helpdesk. Those are convenient if you already live in that helpdesk; standalone vendors usually go deeper on agent quality.
Where it doesn't fit
An AI agent is the wrong tool for three kinds of conversation:
- Anything where the customer needs real empathy. Billing complaints from a long-term customer, account closures, anything where they're already upset. Hand to a human early.
- Anything with a regulatory script that must be read verbatim. Insurance disclosures, certain healthcare flows, KYC steps. The compliance risk of paraphrasing is bigger than the time saved.
- Anything the customer has already escalated. Once a person asks for a human, putting them back on the bot burns trust fast.
How to evaluate one in a pilot
Test on conversations the agent has never seen, not the vendor's sandbox. The questions that matter:
- Resolution, not deflection. Did the customer's actual problem get solved? See resolution rate.
- What happens when it doesn't know. A good agent says "I'm not sure, let me get someone"; a bad one makes something up.
- The handoff to a human. Does the human get the conversation context, or do they restart from zero?
- The action layer. Can the agent actually do refunds, edits, escalations in your systems — or only describe how the customer could do it themselves?
How it relates to the metrics around it
An AI agent improves deflection almost by definition — fewer conversations reach a human. Whether it improves resolution is a separate question. And the answer to that question lives in recontact rate: how many of those "resolved" customers came back about the same thing within 72 hours through another channel.