Done for you, or a platform you configure yourself?
A self-serve platform sells you a console and your own time. A done-for-you build sells you a result. The right answer depends on whether your evenings are worth more than the difference.
What do the cheap AI platforms actually give you?
A working product and an empty configuration. You write what the assistant knows, decide what it refuses, connect it to your systems, and maintain it when a model update changes its behaviour.
None of that is hidden or dishonest. It is what self-serve means, and for somebody technical with time it is a bargain.
Where do self-serve setups usually go wrong?
The refusals. Almost nobody configuring their own assistant sits down and writes the list of questions it must never answer, because the platform does not ask for one and nothing fails until a customer asks the wrong thing.
The second is drift. It gets set up in a good week and nobody revisits it after a model update, so its behaviour changes quietly.
Who should use a platform rather than an agency?
Anybody comfortable writing the knowledge and the rules themselves, with time to maintain it, in a business where a wrong answer is embarrassing rather than dangerous. That is a lot of businesses and they should save the money.