Assist drafts. An agent can finish.
Agent assist is a copilot beside the person who owns the ticket. It suggests a reply, a classification, or a next action. The person accepts, edits, or ignores it, and the desk records that person as the one who acted.
An AI agent is given the step. It can answer the customer, route the ticket, or write a field under rules you set. A person is still in the design of those rules, but not necessarily on every ticket.
Read the control, not the name.
The labels overlap in vendor copy. A product called an agent may still wait for a click. A product called assist may send on its own when a score is high. The name on the pricing page is not the control.
The useful test is the unattended ticket. If nobody is at the desk, does this product finish that ticket, or does it wait? That answer is the product you are buying.
Faster people, or a different doer.
Assist is the right buy when every answer still needs a person, and the goal is to make that person faster. You are staffing the same queue, with a shorter handle on the work they still touch.
An agent is the right buy when a defined slice of the queue should close, or move, without waiting for that person. You are changing who does the work, not only how fast they type.
Neither label updates the record by itself.
A fluent customer reply can still leave the order, the CMDB, or the account record untouched. Ask where the write lands. Finish means the record changed, not only that a message went out.
Some vendors sell both, priced apart. A priced example is the Intercom comparison, because that page carries the source and the date. This page does not restate the rate.