The point where it starts improving itself

Two things, and neither is available to a service bought on its own. The first is one memory: every agent reads from and writes to the same record, so a client explains themselves once rather than once per channel, and a correction you give is written down once rather than repeated to each tool in turn. The second is a loop that reads what actually happened, proposes one improvement, and proves it before it ships, including improvements that mean changing the software. A single service can be good. Only a system that sees all of them can tell you which part of the whole is worst this week.

It writes real code, and that is worth saying plainly because most descriptions of this category promise the opposite. It changes two things: the written layer, meaning prompts, worked examples, routing rules and standing constraints, and the software itself. Which one is being changed decides what grades it. A change to the software goes through a type checker and a test suite that do not care how confident anything sounded. A change to what an agent says is replayed against conversations that already happened and compared against what is currently live. Either way it is one change at a time, it ships only if it wins and only after you approve it, and nothing changes without a record of what changed and why.

In files you can open, not inside a model. There is the record of the conversations, the written instructions your agents run on, and a set of notes about your business that the system has to read before it starts any piece of work. That last one is where a correction goes when you give it: a line in a file, with a date on it, rather than something a model is trusted to have absorbed. It is worth insisting on that shape whoever you buy from, because a memory kept in files is one you can read, correct and take with you, and a memory kept inside somebody else's model is one you are renting.

It learns from your own conversations, which is the only material that could make it better at your business specifically. Where that material is held and what the underlying model vendors may do with it are two separate contract questions, and they are worth asking in writing rather than assuming: that your conversations are stored in an environment you control, that they are not used to train anything general, and that they can be deleted on request. Those are the same three lines worth insisting on with any vendor in this category.

The shared memory helps immediately, because it removes repetition from the first conversation onwards. The improvement loop is slower and honestly so: it needs enough real conversations to tell a genuine weak point from a bad week, which for most brokerages is a matter of months rather than weeks. Anybody quoting a timeline shorter than that is quoting one they have not measured.

It gets undone, and the reason that is a short answer is that the loop is built backwards from it. Every change is one change rather than a batch, it is compared against what was live before it shipped, and the comparison and the approval are both written down. So a change that looks wrong a fortnight later can be identified and reversed rather than debated, which is the difference between a system that learns and a system that drifts.