AI in Buy Now Pay Later: Decisions, Disputes and Affordability

Buy now pay later operates under two pressures that pull against each other: decisions must be instant and frictionless to preserve conversion, and they must increasingly meet consumer credit standards for affordability, explanation, and fair treatment. Automation has to serve both. This guide covers where it helps and where the regulatory boundary sits, drawing on FISTA Solutions' AI agents work in fintech. It complements ai in banking and what is the right to explanation. This article is general guidance, not legal or regulatory advice.

What constrains the decision?

Latency and defensibility together. Sub-second response is a product requirement, since friction at checkout costs conversion directly. And the decision must be explainable and fair enough to survive scrutiny, which is a different kind of constraint.

Those two push toward deterministic scoring with recorded reasons rather than complex models whose outputs cannot be accounted for individually. Speed and explainability happen to be aligned here, which is fortunate.

Element	Automatable	Constraint
Identity and fraud checks	Yes	Latency
Credit and affordability scoring	Rules and models	Explainability, fairness
Adverse decision reasons	Yes, from recorded basis	Must be meaningful
Merchant onboarding checks	Yes	Risk coverage
Dispute and chargeback handling	Largely	Consumer protection
Collections communication	Partly	Vulnerability rules

Why must adverse decisions be explainable?

Because declining a consumer affects them, and in many jurisdictions they are entitled to know why and to contest it. An explanation must be meaningful to the person rather than technically complete — which factors drove the outcome and what would change it.

That requirement shapes the architecture. A system that records the decision basis at the time can explain it; one that reconstructs a plausible explanation afterwards is on weaker ground. See what is the right to explanation.

What is changing in affordability?

Regulators in several markets have moved BNPL within consumer credit frameworks, bringing affordability assessment obligations with them. That converts decisioning from a commercial risk calculation into a regulated process with documentation and fairness requirements attached.

The practical consequence is that the decision record matters as much as the decision, and systems built purely for speed and loss minimisation need retrofitting to produce evidence.

Why is merchant risk under-managed?

Because attention concentrates on the consumer side. Merchant fraud, delivery failure, and insolvency all generate disputes, refunds, and losses, and monitoring merchant behaviour — dispute rates, delivery complaints, sudden volume changes, category drift — is frequently less developed than consumer decisioning.

It is also straightforwardly automatable, since the signals are in the provider's own transaction data.

How are disputes handled?

Largely mechanically. Gathering the order, delivery evidence, merchant response, and consumer statement, applying the scheme rules, and producing an outcome with reasons. High volume, rule-governed, and currently a substantial manual cost.

What needs care is the consumer-facing communication, which must be clear and must not read as dismissive, and the escalation path for consumers who disagree.

What constrains collections?

Consumer protection rules on contact frequency, timing, tone, and the treatment of customers in financial difficulty. Automated sequences that ignore vulnerability signals create real harm and regulatory exposure.

Detecting signals of difficulty and routing those customers to trained humans quickly is more important than optimising contact effectiveness, and it should be built as a hard rule rather than a preference.

What about fairness testing?

Increasingly expected. Decisioning that produces materially different outcomes across protected characteristics is a regulatory problem regardless of whether those characteristics were inputs, because proxies exist. Testing outcomes by segment, documenting the results, and acting on disparities is part of operating the system rather than a research exercise.

How is it evaluated?

Approval and conversion rates, loss rates, dispute rates by merchant, complaint volumes, adverse decisions successfully challenged, and outcome disparities across segments. Conversion alone is the metric most likely to be optimised and the most likely to create problems elsewhere.

What goes wrong?

Decision models whose basis cannot be explained. Affordability treated as a commercial rather than regulated question. Merchant monitoring left to manual review. Collections automation without vulnerability routing. And optimising conversion without watching complaints and losses alongside.

What does it cost to run?

Per decision, very little; volume is the driver. The substantial costs are the evidence and documentation infrastructure the regulatory position requires, and the fairness testing regime, both of which are ongoing and both of which are frequently underestimated at build time.

What should you do first?

Check whether you can produce, for a declined application from last month, a clear statement of why. If that requires investigation rather than a lookup, that gap is the priority regardless of what else is planned.

How does this differ by market?

Considerably, because regulatory treatment of BNPL varies by jurisdiction and is changing in several at once. A decisioning system built for one market's requirements may not satisfy another's, particularly on affordability evidence and adverse decision explanation.

Designing the decision record to capture more than the current market requires is cheaper than retrofitting it when a second market's rules apply, and it is the kind of decision that is inexpensive at build time and expensive later.

How FISTA Solutions helps

FISTA Solutions builds BNPL decisioning and operations with recorded decision bases that support meaningful explanation, affordability evidence captured at decision time, merchant behaviour monitoring from transaction data, dispute handling against scheme rules, and collections with hard vulnerability routing, through AI agents, AI enablement, and forward deployed engineers. The record behind the approach is 150+ projects for 50+ companies with 99.9% uptime.

To build BNPL operations that satisfy regulators and conversion targets, message FISTA on WhatsApp, or read what is the right to explanation.