1. Interfacing with the traditional (fiat) banking system: today this can be done with regulated fiat-backed stablecoins like Paxos USD, Gemini USD or Circle's USDC. It's critical that these institutions be carefully regulated by government entities to ensure the deposits are protected, but this is feasible (Paxos and Gemini are regulated by NYDFS, Circle's regulation story is more complex but there is one.) This regulation is crucial to ensure that "the assets are in the bank", but you only need a few such issuers to be regulated. This is much more scalable than trusting every two-bit crypto exchange in the world.
2. Custody of signing keys: this can also be technically managed by centralized institutions. Anchorage Digital and Fireblocks are examples of this. These centralized solutions can't ever perfectly eliminate risk (they are centralized, after all), but they can make you much resistant to rubber hose attacks. The actual technology is based on hardware security modules and MPC, plus lots of excellent security engineering.
In theory once you have reliable building blocks for these two components, the rest can be achieved using smart contracts without the need for further centralization or trust risk. (You will, of course, risk bugs in the smart contract. But this is a technical problem that we can potentially solve using technical means, not a fundamental problem of human trust.) In general the idea here is to optimize your "trusted human base" so that it is as small as possible (and have those parts carefully regulated), and then you can covert the rest into problems that have purely technical solutions.
Obviously this story only works if you're doing fiat-crypto conversion and storage of assets. The minute you start lending assets to people then you're back to trusting human beings to actually pay you back, so you'd better make good judgements about their (possibly illiquid) collateral.