Most build-vs-buy comparisons start with a cost spreadsheet. That's the wrong first question — the right one is who owns the problem when the AI feature is confidently wrong in front of a customer.
Buying a vendor tool means the failure mode is partly theirs: support tickets, model updates, and liability sit with a company you can escalate to. Building it in-house means every one of those failure modes lands on your own team, indefinitely.
Vendor tools tend to win when the use case is common enough that someone else has already hardened it — document parsing, standard chat support, transcription. In-house builds tend to win when the use case is specific enough to your data or workflow that no vendor is optimizing for it.
Cost only matters as a tiebreaker once the ownership question is answered. Teams that lead with cost end up buying tools they can't maintain, or building features they can't support — either way, for the wrong reason.
A useful test: if the tool disappeared tomorrow, would you rebuild it in-house, or route around it? If the honest answer is "rebuild," you probably should have built it from the start.