AI agents expose a costly blind spot in fraud prevention

Merchants are facing a new and awkward consequence of the rise of agentic commerce: their fraud systems may be rejecting the very customers they are meant to protect.

AI agents expose costly blind spot in fraud prevention

As artificial intelligence tools begin to search, compare and complete purchases on behalf of consumers, a growing number of legitimate transactions risk being mistaken for hostile automated activity.

Chargebacks911, the dispute management and chargeback prevention specialist, argues that this problem is becoming a material revenue risk for retailers, particularly as banks, card networks and ecommerce platforms accelerate work on agent-initiated payments.

Why old fraud models are struggling

For years, fraud prevention has depended on identifying patterns in human behaviour. Device fingerprints, browsing sessions, mouse movements, click paths and authentication journeys have all helped merchants distinguish genuine customers from criminals. Agentic commerce unsettles that model.

AI shopping agents do not behave like conventional shoppers. They can operate at speed, compare multiple offers, navigate sites programmatically and complete transactions with limited visible human input.

In doing so, they may resemble the very bot activity that fraud systems were designed to block.

The problem is sharpening as automated traffic becomes a larger part of the internet economy. Imperva’s 2025 Bad Bot Report estimates that bots now account for 51 per cent of internet traffic, with 37 per cent considered malicious.

For merchants, the challenge is no longer simply identifying bots; it is distinguishing harmful automation from authorised AI-driven purchasing.

False declines could become a strategic problem

Much of the industry debate around agentic commerce has focused on liability: what happens if an AI agent buys something a customer later disputes? That is a legitimate concern. But the opposite scenario may prove more immediate.

If a merchant declines a valid AI-initiated transaction, there is no chargeback to record, no dispute to analyse and no obvious failure point. There is simply lost revenue.

The commercial implications could be wider still. Repeated false declines may weaken customer trust, damage conversion rates and reduce a merchant’s visibility to AI agents that learn where transactions succeed.

In a world where agents may increasingly mediate consumer choice, being “agent-friendly” could become a competitive advantage.

Consent trails become the new evidence layer

Chargebacks911 argues that merchants need to shift from relying solely on real-time behavioural signals towards a more robust record of consent and authority.

The key question is not whether the transaction looks human, but whether the agent was permitted to act, within what limits, and whether the final purchase matched those permissions.

That requires infrastructure capable of capturing a timestamped trail of authorisation, decision-making and execution. Systems such as Chargebacks911’s Unified Dispute Management System and ResolveLab are designed to help merchants build this evidence architecture, using AI and machine learning to analyse permission frameworks alongside transaction data.

Merchants need to adapt quickly

Agentic commerce is moving from experiment to infrastructure. Visa, Mastercard and major ecommerce platforms are already testing or enabling forms of AI-assisted transaction. Merchants that continue to treat all unusual automated behaviour as suspicious may protect themselves from some fraud, but at the cost of rejecting future revenue.

The next phase of payments will demand a subtler distinction: not human versus bot, but authorised agent versus malicious automation. That distinction may soon define the difference between merchants that participate fully in AI-led commerce and those quietly excluded by their own defences.