A layer finishes work in place.
An AI support layer reads the ticket your team already has, decides what should happen, and writes the result back into the helpdesk or ITSM that holds the record. The desk stays the system of record.
The point of the layer is the update behind the reply: a classification, a drafted resolution, a field written in the system the rest of the company already trusts. A conversation that ends somewhere else has not finished the job.
It is not a rip-out.
Ripping out a helpdesk means moving queues, history, SLAs, and every integration hung off them. That is a platform project. A layer assumes the desk stays and does its work inside the instance you already license.
A chat window in front of the queue is a different product again. It can deflect a contact. It does not, by itself, close the ticket or update the systems the ticket is about.
Finish means the record changed.
Vendors disagree about what a finished job is. Some count a conversation closed in their own channel. Others count an assisted reply inside the desk. A layer is aimed at the record: the ticket, and the system behind it, show what was done.
That difference is why a buyer comparing vendors should ask where the work ends, not only how fluent the reply sounds. The compare pages on this site take that question vendor by vendor, with dates on the sources.
A rip-out is sometimes right.
Leave the helpdesk when the desk itself is the problem. The product is ending, the contract will not renew, or the team has already chosen the next system of record. An AI project will not rescue a platform you have decided to exit.
If the desk is staying, the useful question is which AI should sit on it: the one the vendor bundles, or a layer that can also reach the systems around the desk. That choice is the next guide.