1) Encryption 2) Date/time math 3) Distributed locking/concurrency primitives
All of these are of sufficient complexity to virtually guarantee that you’re going to miss a score of subtle edge cases, and the consequences thereof are likely simultaneously pervasive, silent and catastrophic.
I mean, obviously, someone has to code crypto, or else it wouldn't exist. But, whenever you try talking about the topic online -- whether you're doing it professionally or for self-education -- a probably well-meaning chorus of spam comments warn against doing anything.
I suspect that some crypto-folks have given up rebutting the don't-roll-your-own stuff and just down-vote reflexively.
I still think that ties into the original question, when does it make sense to roll your own? If it's not part of your core business and doesn't give you an edge, then it never makes sense, unless you're doing it for the fun of it.
You have millions in the project and a dozen PhDs working with you, and you've already authored public papers demonstrating the weaknesses of things previously believed to be strong.
If you are not at least one of those things, it's malpractice for you to even consider letting your employer pay for your crypto "invention"s. That's like a nurse building her own OR and insisting the hospital let her cut people open in it. No. Don't. Stop.