Banks found it relatively easy to fend off the crypto bros. Artificial intelligence agents may prove a lot harder to tame.

This time last year, stablecoins were all the rage. Armed for the first time with legitimacy from the US Genius Act, these 1:1 digital clones of the greenback seemed poised to go mainstream and take away some of the $19 trillion parked with US lenders — and another $14.5 trillion in dollar deposits at overseas banks. Or so went the story.

A year later, the revolution seems to have fizzled out. Circle Internet Group Inc., the issuer of USD Coin, is valued at $22 billion, down three-fifths from its June 2025 peak. Coinbase Global Inc., the exchange for these tokens, is seeing more institutional adoption, as the likes of Bank of New York Mellon Corp. and DBS Group Holdings Ltd. seek partners to handle demand for crypto services. But stablecoins are a long way from replacing the bank account. For the foreseeable future, they are more likely to be used by corporate treasurers to move funds to countries with poor financial infrastructure than by shoppers standing in a convenience-store checkout.

Banks, however, can’t rest easy. While they were busy lobbying to protect deposits from crypto, AI got smarter. The technology may now be good enough to shake loose the deposits that are just sitting in a client’s account. The days when lenders made easy money from customer inertia are over.

Transactions where no human swipes a card, presses a button, or scans a QR code aren’t exactly new. As McKinsey & Co. points out in its , banks are used to such "invisible" payments where the customer remains in control. Subscriptions and standing orders have been around for decades, while printers reordering ink and connected cars paying tolls have expanded the scope of automation. The crucial difference is that the rules for these are set by humans well in advance.

The real threat comes from autonomous AI agents that make financial choices at the exact moment of execution. Rather than requiring customers to specify whom to pay, when and how much, these programs operate under broad mandates, weighing yields, fees and liquidity in real time. As the technology matures, households will increasingly trust personal assistants like Meta Muse to buy groceries wherever they are cheapest — using whichever card offers the best rewards. Companies, meanwhile, could source from suppliers offering the best combination of prices and credit terms.

In Western markets, Visa Inc. and Mastercard Inc. are racing to establish rules for verifying agent identities, while Google and OpenAI are writing protocols that allow AI to transact without interacting with conventional checkout websites. But the agents themselves may prefer low-cost, state-backed payment networks — such as Brazil’s Pix or India’s UPI — which are far easier to navigate.

It won’t be just spending habits that will change. Saving patterns will, too. Customers who once left cash languishing in low-interest checking accounts will increasingly rely on AI assistants to hunt for better yields at lower cost. If an idle dollar can earn even slightly more across the street, an AI tool can sweep it there in milliseconds.

That could get expensive for banks. McKinsey estimates that AI agents could put $75 billion to $160 billion of global payments revenue at risk. Unsurprisingly, the biggest pot is the money lenders make from customer stickiness. In the consultancy firm’s base case, $65 billion of net interest income on retail and commercial deposits is at risk.

The third-party vendors controlling these agent mandates would turn transaction banking on its head. Instead of corporate treasury chiefs logging into a bank portal, independent platforms like Kyriba, Trovata, GTreasury, and Agicap — or enterprise software behemoths like SAP SE and Oracle Corp. — could deploy autonomous tools directly in financial workflows. Software vendors would hold the client relationship and dictate where to route funds, reducing traditional lenders to dumb warehouses for account balances.

The writing on the wall is clear: If banks provide the agents that advise corporate clients and execute their mandates, they can preserve both the relationship and the deposits. Cede that role to technology vendors, and they risk losing both.

At the same time, deposit-taking institutions must spend heavily on security. AI assistants carry digital identities — linked to their owners’ electronic devices — that lenders need to authenticate in real time. And should an agent behave suspiciously in its relentless pursuit of yield, compliance systems will need split-second accuracy to stop transactions before they clear.

All this will be expensive, and the bill will come due just as income from those idle deposits shrinks. That’s the real challenge. By comparison, the threat from crypto exchanges luring customers with manufactured yields on zero-interest stablecoins looks modest. AI is the technology banks should worry about.

For agile banks, however, there is a silver lining. As Accenture notes in its report, generative and agentic AI are driving down operational costs by automating software engineering and compliance. The winners will reinvest those savings to build the tools that keep clients in the fold. The laggards will face rising investment costs just as the margins they earn on deposits are being squeezed.