This is part one of Sandmark's nine-part series on Agentic AI.
For years, banking apps were the main interface between customers and their banks. That's starting to change: instead of people initiating transfers themselves, AI agents are increasingly doing it for them, autonomously.
An AI agent is a software system that uses artificial intelligence to pursue goals with a degree of autonomy. Unlike a standard chatbot exchange, an agent can plan multistep workflows, pull in live data and operate external tools, analyzing market conditions, weighing risk and executing transactions, without a human signing off on every step.
For consumers, the appeal is fairly straightforward: an agent that can pay your bills, shuffle your savings or manage your investments without you having to lift a finger. For institutions, the questions are different, and that's where things get more complicated.
AI agents and crypto, or Web3, the blockchain-based alternative to today's centralized internet services, fit together unusually well. Traditional financial systems are layered with intermediaries, manual checks and legacy APIs that autonomous software struggles to navigate. Blockchains, shared digital ledgers that record transactions across a network of computers rather than one central server, offer something closer to a native environment for it: open, machine-readable and not dependent on a central gatekeeper approving each step.
Transactions are recorded in cryptographically linked blocks that are very resistant to tampering: a new block is only added once a majority of independent nodes, the individual computers that make up the network, agree it's valid. That structure doesn't require a trusted intermediary, which suits software that's meant to act on its own.
Smart contracts add another layer: self-executing code that carries out a predefined action once conditions are met ("if payment X is received, send token Y"), without a human in the loop. This is what underpins DeFi (lending, trading, staking) and NFTs (automated royalty payments).
Just as important is composability, crypto protocols are built to plug into each other, so an agent can move across lending platforms, exchanges and wallets in a single workflow. TradFi systems, by contrast, are far more siloed; getting one institution's systems to talk to others is often a project in itself. That interoperability is a big part of why agents can operate so fluidly in crypto and not (yet) in traditional finance.
For institutions, the interesting questions aren't really about convenience, they're about control. Who gives final approval on a transaction? Who's liable if an agent gets it wrong? How do you audit a decision-making process that isn't fully transparent, even to its own developers? We'll dig into these questions in more depth in other articles in this series.
The risks aren't hypothetical. Agents can be manipulated, through prompt injection, hidden instructions smuggled into the data an agent processes to hijack its behaviour, or tampered code, for instance, into routing funds to an address the attacker controls, effectively laundering the theft through a "legitimate" automated transaction. Add to that the general opacity of how these models reason, and the picture becomes one institutions will need to take seriously before handing over the keys.
