Every major payments conference this year has featured some version of the same conversation: AI is transforming fraud detection, Machine learning models are getting better at identifying suspicious transactions in real time, the checkout is becoming smarter, faster and more secure. While all of that is true, almost none of it addresses where trust in digital commerce is actually breaking down.
Commerce has already moved beyond the checkout moment
Consider how much of digital commerce no longer involves a human being at the point of purchase. Subscriptions charge automatically, platforms execute repeat purchases based on saved preferences, algorithms reorder inventory and increasingly, AI agents are being granted delegated authority to make purchasing decisions on behalf of consumers without any human involvement at the moment of sale.
That all means that the transaction happens, the authorisation succeeds but the dispute begins later. And addressing that later moment requires something the industry has largely overlooked: intelligence that operates after the payment has been approved.
The assumption embedded in most AI investment
The assumption embedded in most AI fraud investment is that the critical moment is the transaction itself. Stop the bad actor at the point of purchase and the problem is solved. It is a logical place to start and the technology has improved. But it reflects a mental model of commerce that is becoming increasingly outdated, one in which a human being consciously initiates every purchase, is present at the moment it occurs and can be held accountable for what happens next.
Most AI deployment in payments has therefore focused on pre-transaction signals: behavioural biometrics, device fingerprinting, velocity checks, anomaly detection at the point of authorisation. These are valuable capabilities but they generate evidence about what happened before a transaction was approved, not about what happened after. And it is the after that determines whether a dispute can be fairly arbitrated.
We have spent years teaching AI how to approve payments but I argue that we’ve barely started teaching it how to defend them.
The gap between investment and outcomes
Research published in the 2026 Chargeback Field Report, based on survey data from more than 250 merchants, found that whilst over a quarter of merchants already use AI-based fraud prevention tools and another 37% plan to adopt them, the average net revenue recovery rate from disputed transactions sits at just 10.7%. That means merchants are investing in AI but the outcomes have not followed.
The reason, I believe, is structural. When a consumer initiates a chargeback, the question being asked is not whether the authorisation was technically valid. It is whether the purchase was genuinely intended, whether the goods or service were delivered as promised and whether the merchant can demonstrate that. Those are questions about evidence, accountability and intent and not questions that pre-transaction AI was designed to answer.
The result is a growing asymmetry – merchants have invested heavily in the systems that approve transactions and comparatively little in the systems that defend them. If you cannot accurately identify which disputes are fraudulent, you cannot represent them effectively. If you cannot represent them effectively, you cannot recover the revenue. The questions that matter in a dispute – evidence, accountability and intent – sit outside the scope of what pre-transaction AI was built to address.
What post-transaction intelligence actually requires
I think of this as post-transaction intelligence: a discipline distinct from fraud detection, one that needs to work across fragmented evidence sources, such as payment gateways, customer service records, delivery confirmations, communication logs, behavioural data from the full transaction lifecycle. It needs to understand context, construct a coherent account grounded in evidence, accountability and intent not simply whether the transaction should have been approved in the first place.
That is a harder problem than fraud detection at checkout. It is also a more consequential one, because it is where the financial damage is actually occurring and where the trust that sustains long-term customer relationships is ultimately won or lost.
The capability to build post-transaction intelligence already exists. Specialist dispute management platforms are already doing this work – connecting fragmented evidence sources, applying machine learning to identify patterns across the full transaction lifecycle and translating that insight into measurable dispute outcomes. What has been missing is the industry’s willingness to treat the infrastructure as a strategic priority.
The harder question AI must answer
The merchants who understand this shift first will have better evidence, better customer conversations and, ultimately, stronger trust. For years the industry has asked whether AI can decide who should be allowed to buy but surely the harder question is whether it can explain those decisions when trust is challenged afterwards.
