BNPL Loan Values Are Rising—AI Can Keep Risk in Check

The average yearly dollar value of a buy now, pay later (BNPL) loan climbed to $848 in 2023, up from $745 in 2022—a 14% increase based on analysis released by the Consumer Financial Protection Bureau (CFPB). That’s not a trivial shift. Bigger average loans change everything: underwriting risk, fraud incentives, repayment behavior, and the pressure on payments infrastructure that has to approve, route, and settle these transactions in real time.

Here’s the uncomfortable truth: most BNPL stacks were built for growth first and control second. When average loan size rises, that tradeoff gets exposed. And as we head into year-end spending patterns (returns, gift-card abuse, account takeovers, first-party fraud), BNPL becomes even more attractive to bad actors because it sits at the intersection of instant approval and delayed repayment.

This post is part of our AI in Payments & Fintech Infrastructure series, and I’m going to take a clear stance: BNPL growth isn’t the problem—weak, slow risk systems are. The good news is that AI (used correctly) can close the gap without wrecking conversion.

What the CFPB numbers really signal about BNPL risk

BNPL is getting larger per loan, and that changes the economics of every decision engine.

The CFPB’s study (covering 2022–2023 and summary data from six major BNPL providers representing roughly 40% of the point-of-sale financing market in 2023) highlights three shifts that matter for operators:

• Average BNPL loan value increased 14% to $848 (2023 vs. 2022).
• Total BNPL originations grew 23% year over year (26% inflation-adjusted), but growth is slowing relative to earlier years.
• Operational loss indicators improved in the dataset:
  • Late-fee incidence fell to 4.1% (down from 5.2%).
  • Charge-offs declined to 1.83% (down from 2.63%).

Those improvements sound reassuring, and they may be real inside that sample. But there’s a catch that risk leaders shouldn’t ignore: late fees and charge-offs are lagging indicators. They tell you what happened after you already approved the loan.

If your main comfort metric is “charge-offs are low,” you’re measuring the past—not controlling the present.

This is also where the public debate gets messy. Consumer advocates argue that repayment stress isn’t fully captured by late-fee and charge-off stats because:

• Many BNPL products rely heavily on automatic payments, which can trigger overdraft fees even when the BNPL loan itself is “paid on time.”
• BNPL payments can create cash-flow shortfalls that don’t show up as delinquency.

Meanwhile, industry groups point out (fairly) that transaction data beats surveys for precision—yet surveys can capture consumer experience that raw repayment data won’t.

The operational takeaway is simple: BNPL risk isn’t one metric. It’s a system behavior. And systems need real-time controls.

Why rising BNPL loan sizes attract smarter fraud

Higher average loan values increase fraud incentives because the payoff per successful attempt rises.

When a typical BNPL ticket grows, the fraud math changes for attackers:

• More value per checkout means fewer attempts are needed to generate the same profit.
• Attackers can concentrate on high-value merchants, categories, or gift-heavy seasonal spikes.
• Fraud can blend into legitimate demand because “bigger baskets” aren’t automatically suspicious.

The fraud patterns BNPL operators see when amounts rise

Rising BNPL loan values tend to amplify a few recurring patterns:

1. Account takeover (ATO) + BNPL checkout: Attackers use compromised credentials and pick BNPL to reduce immediate friction.
2. Synthetic identity at scale: A synthetic profile can “warm up” with small loans, then step up to larger baskets.
3. First-party fraud / friendly fraud: A borrower intends not to repay, or disputes the transaction after goods are received.
4. Refund and returns abuse: A borrower gets refunded to a different instrument or exploits return windows.

What makes BNPL unique is that fraud risk is split across multiple parties—merchant, BNPL provider, payment processor, sometimes a bank partner—and each party sees only part of the picture.

This is exactly where AI belongs: connecting partial signals into a single risk decision fast enough to matter.

Where AI fits in BNPL: risk, fraud, and routing (the practical version)

AI improves BNPL outcomes when it’s used to make better real-time decisions, not when it’s used as a fancy reporting layer.

If you’re building or operating BNPL infrastructure, the three highest-ROI AI applications are:

1. Real-time fraud detection (stop bad checkouts without blocking good ones)
2. Dynamic credit risk management (approve responsibly without crushing conversion)
3. Intelligent transaction routing (raise auth rates and reduce processing cost)

AI for BNPL fraud detection at checkout

The best BNPL fraud systems treat checkout like a high-speed investigation.

Effective AI models typically combine:

• Device and network signals (device fingerprint stability, IP reputation, velocity)
• Behavioral signals (typing cadence, session flow anomalies, bot-like patterns)
• Identity signals (email/phone tenure, mismatch patterns, KYC confidence)
• Merchant context (category risk, return rates, AOV distribution shifts)
• Payment signals (AVS/CVV results where available, token history, decline reasons)

A key stance I’ll defend: fraud decisions should be explainable enough to operate, even if the model is complex. If your ops team can’t tell why approvals dropped 3% on Tuesday, you don’t have a risk engine—you have a black box.

Practical outputs to build toward:

• A single risk score plus 3–5 human-readable reason codes
• Step-up paths (e.g., one-time passcode, bank account verification, stronger identity proofing)
• Real-time velocity controls (per user, per device, per merchant, per BIN range)

AI for underwriting: move from static rules to dynamic affordability

When BNPL loan values rise, static rulebooks get brittle.

Traditional underwriting often looks like:

• “Approve if score > X”
• “Decline if too many recent loans”
• “Limit amount by segment”

That approach breaks when borrowers’ cash-flow patterns are volatile (which is common) and when macro conditions are uncertain.

AI-based underwriting works better when it predicts repayment probability conditional on current context, such as:

• recent repayment behavior across BNPL loans (not just lifetime)
• basket composition (some categories correlate strongly with returns and disputes)
• seasonality (holiday spikes, post-holiday return windows, tax-season dynamics)
• payment method reliability (and how often autopay fails)

If you’re cautious about “AI in credit,” good. You should be. The target isn’t “approve more.” It’s approve the right people at the right limit.

A clean operating principle:

BNPL underwriting should be limit-centric, not approve/decline-centric.

That means AI should frequently answer: “What’s the maximum safe amount for this user right now?”

AI for transaction routing: approvals and cost are connected

BNPL isn’t just a lending decision—it’s a payments decision.

Once approved, BNPL still depends on payment rails and processing steps that can fail or cost more than expected. AI-driven routing matters because:

• routing can lift authorization rates by choosing the best path based on issuer behavior
• routing can reduce cost by selecting processors or rails that perform best for that merchant/user profile
• routing can minimize operational risk by avoiding paths correlated with chargebacks, retries, or reconciliation noise

For payment platforms, this is a big deal: a 50–100 basis point auth-rate swing at scale is not a rounding error. It’s revenue, merchant retention, and risk exposure.

The metric trap: why “late fees are down” isn’t enough

Late fees and charge-offs are useful, but they don’t tell you whether your system is creating hidden consumer harm—or hidden portfolio risk.

If you’re responsible for BNPL risk, add these operational metrics to your dashboard:

• Autopay failure rate (and how often failures recover after retries)
• NSF/overdraft proxy signals (where permissible and available)
• Return-adjusted loss rate (fraud and disputes often hide in returns)
• Repeat BNPL stacking (multiple concurrent plans across merchants)
• Time-to-first-miss (how quickly new cohorts miss a payment)
• Reasoned declines (declines by reason code to spot model drift)

This matters because regulatory scrutiny is shifting from “are people paying late?” to “are people being harmed even when they pay?” That distinction will shape product design and model governance.

A practical AI playbook for BNPL lenders and platforms (next 60 days)

You can improve BNPL risk controls quickly if you focus on the decision points that actually move outcomes.

Here’s what I’d implement first if the goal is better control without tanking conversion.

1) Map your real-time decision journey

Document the exact flow from checkout to settlement:

• identity inputs
• fraud checks
• underwriting/limit decision
• step-up verification
• payment authorization
• fulfillment signals
• servicing and collections signals

Most teams discover they have duplicate checks in the wrong order—or missing data at the moment it’s needed.

2) Add step-up verification instead of hard declines

Hard declines are expensive when you’re wrong.

Use AI to trigger step-ups only when needed:

• OTP or passkey confirmation
• bank account verification for higher limits
• stronger KYC for suspicious patterns

3) Build a model monitoring routine that ops can run

Model drift is guaranteed in BNPL because consumer behavior changes by season and merchants change their promos.

At minimum:

• weekly cohort performance review
• alerts on approval rate swings and loss-rate spikes
• a “challenger model” running quietly to compare outcomes

4) Connect fraud + credit + payments data (even if you can’t centralize it)

You don’t need a perfect data lake to get value.

Start by making key signals shareable in near-real time:

• device and identity risk outputs into underwriting
• underwriting limit outputs into routing
• disputes/returns into fraud training

BNPL is an end-to-end system. Treating it as three separate teams will keep creating blind spots.

What BNPL growth means for fintech infrastructure in 2026

Rising BNPL loan values are a stress test for modern payments infrastructure.

As the market matures, winners will look less like “fast approvers” and more like “reliable operators”:

• higher-quality approvals (limits that match affordability)
• fewer fraud losses without punishing good users
• cleaner reconciliation, fewer disputes, better servicing
• routing strategies that protect auth rates and margins

And regulation will keep hovering over the sector, especially as oversight shifts and states push for more visibility into who uses BNPL and how repayment plays out.

From the perspective of this AI in Payments & Fintech Infrastructure series, the direction is clear: AI isn’t an add-on for BNPL anymore. It’s the control plane.

If you’re building BNPL, supporting it as a processor, or enabling it as a platform, the best next step is a focused assessment: where do decisions happen, what signals are missing, and which AI interventions can reduce risk today without creating a customer experience tax.

The question I’d leave you with: when the average BNPL loan climbs again, will your risk system get smarter automatically—or will you be back in the war room rewriting rules?