As AI agents grow in popularity and scope, experts have raised questions about responsibility for loss when an agent autonomously makes a payment, with valid credentials, without its owner’s intent or approval.
The topic arises as Nigeria’s digital payments ecosystem expands, and exposes potential gaps in authentication, fraud detection and liability rules.
Rodion Sorokin, chief technology officer at The Capital Index, said conventional fraud controls struggle to distinguish legitimate agents from compromised systems using valid credentials. “To the bank, it looks like an authorised payment. The agent used the customer's valid credentials, so every check passes,” he said.
However, malicious instructions hidden in emails or webpages could manipulate an agent into making payments its owner never intended, said Raj Kamal, founder and chief executive of payments company TransFi.
“Letting an AI agent pay for you is like signing a power of attorney, and most payment systems can't read the terms yet,” Kamal said.
Allan Dabre, senior manager for technology compliance and AI risk at PwC, said pre-deployment testing often assesses whether agents complete tasks, without adequately testing how they respond to manipulation. “Nobody evaluates the failure mode where the agent does the wrong thing successfully,” he said.
The risks raise difficult questions about liability among customers, developers, businesses and payment providers.
Michael Heinrich, chief executive and co-founder of 0G Labs, argued that responsibility should follow each party's control over the transaction.
“The developer owns the agent's behaviour,” he said. “The bank or payment provider owns enforcement: if it agrees to hold the agent to a spending limit and let a payment through anyway, that is on the provider.”
Nigeria's regulatory framework has begun addressing related risks. In November 2025, the Central Bank of Nigeria issued an exposure draft on authorised push payment fraud. However, the proposal does not specifically resolve liability when an AI agent exceeds its owner's instructions.
“Payment rules around the world assume a human starts the transaction,” Heinrich said.
The experts called for individual agent identities, explicit spending limits, verifiable records of user instructions and mechanisms to revoke access quickly. Payment networks are also developing infrastructure for agent-driven commerce, although these initiatives do not settle liability disputes.
For Nigerian financial institutions, the challenge is to establish not just whether a payment used valid credentials, but whether the agent had authority to make it, and who must reimburse the customer when that authority is exceeded.
Share
