The Problem with Autonomous Spending
As AI agents become capable of executing financial transactions, a fundamental problem emerges: following a budget does not mean understanding the job. Daniel Ishag, chief commercial officer at Bakkt, illustrates the concern with an example of an AI agent booking a business-class flight to the wrong destination—potentially a critical error despite staying within approved spending limits.
The issue extends beyond simple misunderstandings. Agents can be manipulated by malicious instructions, accounts can be compromised, or software bugs can trigger unintended payments. Each failure produces the same result—an unauthorized transaction—but raises different questions about who bears responsibility.
Current Safeguards and Their Limits
BeInCrypto's State of AI Agent Payments 2026 found that 14 of 19 reviewed products and workflows could act within pre-approved permissions. However, this capability does not guarantee controls cannot fail. The report examined two incidents in which $281,200 was sent to attackers before recovery occurred, demonstrating that malicious instructions can reach funded wallets without triggering blockchain-level protections.
Payment service terms currently place much of the burden on users. Stripe's Link wallet terms, for example, hold users responsible for transactions initiated by connected agents—even unintended ones resulting from bugs or misunderstood instructions—with exceptions only for Stripe's own technical failures.
Assigning Responsibility
Approving a $200 spend limits what a payment provider knows about a transaction. It may say little about the intended product, recipient, or conditions. After a purchase goes wrong, this gap can become central to disputes over reimbursement.
Some companies are addressing this challenge. Sailo Technologies, a blockchain security firm, offers reimbursement as part of its service. Google's Agent Payments Protocol proposes signed records connecting user instructions, approved purchases, and payments, potentially helping establish what was authorized—though such records do not themselves guarantee reimbursement.
Regulatory bodies are also setting expectations. The UK's Financial Conduct Authority states that regulated firms remain responsible for consumer outcomes when using AI, warning that human review can become merely procedural.
Key Questions for Users
For anyone funding an AI agent, specific questions matter: Which purchases can it make? What record preserves your original instructions? Can you stop further spending? And who handles a payment you challenge?
As automation expands, trust will ultimately be tested after mistakes occur—when an agent makes an error and someone must decide who bears the financial cost.
