Dragonfly Partner Rob Hadick: Agents More Likely to Recommend Mercury
Rob Hadick, a general partner at crypto fund Dragonfly, stated on the MTS program that AI agents may select banks for users. He mentioned that if asked today which bank is agent-friendly, the answer would be Mercury, not JPMorgan.
His judgment has two layers. The first layer denies existing data interfaces: things built on Plaid are unlikely to regularly and massively move any funds. Plaid connects account data and authorization, not the accounts themselves. The second layer concerns the form of banks. He believes that agent-first banks are more likely to emerge, controlled by a financial institution that manages different accounts and can profit from them, and that this institution itself is AI-first and agent-first.
Mercury is a fintech account aimed at companies, offering an official custodial MCP interface that connects compatible AI tools via OAuth. The official implementation is read-only, allowing users to view balances, transactions, cards, and recipients, but assistants cannot directly initiate transfers. Plaid's MCP is aimed at financial data and integration debugging, not banks that hold accounts. Hadick did not say that Mercury can already move money for users, only that the model would first mention it in recommendations.
JPMorgan remains a legacy player in corporate accounts, payments, and custody. Hadick's comparison is not based on asset size, but on whether agents can complete account openings, inquiries, and subsequent fund actions in conversations. He has previously stated that the scale of stablecoin payments could grow tenfold, with value shifting to companies that control distribution, compliance, and settlement, rather than relying solely on the interest from reserves.
This is a judgment on distribution rights, not a new banking license. The recommenders are large models and agents, while the recommended are companies that have made their interfaces agent-readable, bypassing large banks that only exist in human online banking. If funds follow his path, they will flow from the user's main account to accounts that agents can directly access. Benefiting are fintech companies that first hand over read-only interfaces and account opening processes to models; under pressure are middleware that keep interfaces at the Plaid aggregation layer but cannot control accounts and fees, as well as large banks that have not yet become the model's default answer.
Source: Public Information
ABAB AI Insight
Plaid, since around 2013, has sold balances and transactions from thousands of institutions to applications through screen scraping and later bank interfaces. Visa announced an acquisition in 2020, which was terminated due to opposition from the Justice Department. Its product is connectivity, not a balance sheet. Mercury takes a different path: it manages the main accounts, cards, and recipients for startups, keeping the data on its side. Hadick separates the two, implying that agents are moving deposits, not just purchasing another layer of data pipeline.
Dragonfly's investment path is in payment and stablecoin infrastructure, not in U.S. retail banking. Hadick publicly places the next phase of stablecoins in distribution, compliance, and merchant settlement, believing that ledger settlement will compress issuers, acquirers, and card organizations. The agent's choice of bank is an entry point of the same logic: the model first decides where users place their operating accounts, and the account holder then collects deposit interest spreads, card, and transfer fees. Mercury's official interface is currently read-only, indicating that recommendations have occurred, but fund movement has not yet been handed over to agents.
A comparable example is Robinhood, which uses a mobile interface to take retail orders away from Charles Schwab and Fidelity's manual channels; a bank license is not its starting point. Revolut and Nubank also first occupy the application entry, then supplement accounts and licenses. JPMorgan has its own payment and corporate interfaces, but the default answer is not in the conversation. The current stage is a struggle for distribution rights, not a completed deposit war.
Structurally, this represents a transfer of pricing power. The mechanism shifts account opening decisions from customer managers and comparison websites to the models users have already opened. Whoever becomes the default answer for "agent-friendly" first will secure the accounts; Plaid can only read after authorization and cannot keep deposits on its own balance sheet. A read-only interface is a transition: the model is first allowed to view accounts, and the next step is to initiate transfers under limits and human confirmation. The large banks' balance sheets remain, but the recommendation framework is no longer with them.
ABAB News · Cognitive Laws
- Agents first choose the entry point; deposits then decide to stay or leave.
- Data pipelines can connect accounts but cannot retain accounts.
- Read-only is a recommendation; the ability to transfer is banking.