According to CoinMeta, AI wallet proxies need to establish spending policies, rather than just optimizing prompts. Most AI wallet demonstrations focus on whether the proxies can perform exchanges, bridging, staking, and rebalancing independently, which are relatively simple tasks. However, the real challenges arise when proxies encounter contaminated web pages, misunderstand tokens, fail to retry transactions, or follow improper instructions. Normal wallets require approval for each transaction, whereas proxy wallets may need policy support before executing any action. Proxies should be aware of the maximum amount that can be transferred daily and should have access to a whitelist of chains and contracts. There should be strict limits set for slippage, fees, bridging, and token approvals, and certain operations will always require manual signature. Each active permission should come with an immediate emergency stop switch. Before taking action, proxies should also provide a clear and concise receipt that explains what they have seen, their assumptions, planned expenditures, and the conditions for stopping. Prompts are not a sufficient security barrier, as credible websites may alter the models' beliefs. Wallet policies must be kept separate from the models and should be disabled in case of failure. The key question is not whether AI can manage the wallet, but rather how much authority it should be granted before it gains trust.