What problem does it solve?
Buy now pay later works only if the approval decision happens in the time it takes a shopper to reach a checkout button: too slow and the sale is lost, too permissive and losses erode the thin margin on an unsecured, often fee free loan. Unlike a credit card, there is frequently no prior relationship and no traditional credit history to lean on, including for thin file consumers with limited or no credit history.
Reviewing a meaningful share of checkout decisions by hand is not feasible at buy now pay later's volume, so the underwriting model is the business: every basis point of loss it misses is a basis point off a margin that is already thin, and every good customer it declines is a sale the merchant loses and a customer who may not come back.
The models keep having to improve, not just run: repayment outcomes on a high volume, short term loan book arrive quickly, which is both the risk (a bad model season shows up fast) and the opportunity (there is a fast, large feedback loop to retrain on).
How does it work?
- Capture the order and applicant. At checkout, the provider receives the order amount, merchant and available consumer data (device, contact details, and, if returning, prior repayment history on the platform).
- Score in real time. A machine learning model scores the consumer and the specific order using bureau data where available, plus transaction and behavioral signals, without waiting for a manual credit check.
- Decide and size the line. The model outputs an approve, decline or step down (a smaller amount or a different payment schedule) decision and, for returning customers, an available spending limit, in the time the checkout page takes to load.
- Monitor the back book continuously. Delinquency and charge off rates are tracked by origination cohort, so a model or a segment that is underperforming shows up in weeks, not at the next annual review.
- Retrain on new repayment data. As loans mature, their outcomes feed back into the next model generation, so risk separation improves as more repayment history accumulates.
- Audience
- Back office
- Autonomy
- Autonomous
- Adoption
- Mainstream
- Channels
- API and system to system, Mobile app
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Conversion uplift | Too few to pool | 3.4% | 1 | 1 organization |
Value drivers: Risk and loss reduction, Revenue growth, Speed and cycle time, Inclusion and access.
Indicative value
A BNPL provider processing 1 million checkout decisions a month
USD 120,000 to USD 4.1 million
Additional annual revenue from more completed purchases per year
How this is calculated
Formula: decisionsPerMonth * 12 * completionRate * additionalCompletedPurchaseUplift * averageOrderValue * takeRate. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Checkout credit decisions per month decisionsPerMonth, decisions per month | 1,000,000 | 1,000,000 | The reference provider. |
| Checkout decisions that already result in a completed purchase under the prior model completionRate, fraction of decisions | 0.5 | 0.8 | Editorial assumption for the share of checkout decisions that already convert to a completed purchase under the prior underwriting system. Replace with your own approval or completion rate. |
| Additional completed purchases versus the prior underwriting system additionalCompletedPurchaseUplift, relative uplift, not a share of decisions | 0.01 | 0.034 | Conservative against Affirm's reported 3.4% more completed purchases from its transformer based underwriting model, a relative uplift measured against a control group at checkout, not a share of all checkout decisions. |
| Average order value averageOrderValue, USD | 100 | 250 | Editorial assumption for a general merchandise BNPL book. Replace with your own average order value. |
| Net revenue as a share of order value takeRate, fraction of order value | 0.02 | 0.05 | Editorial assumption for merchant and consumer fee revenue net of funding and loss cost. Replace with your own take rate. |
What it leaves out: Revenue only, from the additional completed purchases at an assumed take rate. It leaves out the credit loss on the additional volume, funding cost, and any change in merchant mix or repeat purchase behavior.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Affirm Holdings
United States · Payments and cards · 2026
Affirm, a buy now pay later provider, has underwritten every purchase individually in real time using in house machine learning models for 14 years. In September 2026 it announced a transformer based model that reads the order and timing of events in a consumer's credit history, live at checkout in the US. In its initial deployment the model approved additional eligible applications, including consumers with limited credit histories and no FICO scores, that Affirm's prior models would have declined.
- Conversion uplift: 3.4%
"Measured against a control group, that produced 3.4% more completed purchases, and those additional loans performed better than a comparable expansion under Affirm's previous machine learning models."
Claimed by: organization
Sezzle
United States · Payments and cards · 2025
Sezzle, a buy now pay later provider, discloses in its fiscal year 2025 Form 10-K that its platform reviews the transaction and consumer profile in real time at checkout and that its underwriting platform, informed by its proprietary credit risk models, tailors the lending amount for each consumer. For the resulting receivables portfolio, Sezzle grades credit quality with an internal, proprietary machine learning score it calls the Prophet Score, built from internal risk indicators and consumer attributes predictive of a customer's ability and willingness to repay. Receivables are grouped into three Prophet Score bands, A to C, and Sezzle's risk and fraud team reviews the model's integrity at least annually; the model was last updated in October 2023.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Bureau data where available, plus permissioned transaction and device data
- A growing base of the provider's own repayment outcomes to train and validate against
- Adverse action reason codes mapped to the model's decision factors
Systems to integrate
- Merchant checkout and payment integration (web and app)
- Credit bureau and identity verification services
- Loan servicing and collections systems
- Model monitoring and MLOps pipeline for retraining and champion challenger testing
Complexity: High
The modeling itself is specialized (real time inference at checkout, thin file and no file consumers, explainability requirements for adverse action), and the business depends on continuously retraining as new repayment outcomes arrive, which needs a mature MLOps and model risk governance practice, not a one off build.
- 1
Separate the score from the policy
Keep the statistical model that scores risk separate from the business rules that turn a score into an approve, decline or step down decision, so policy can change without retraining the model.
- 2
Build the adverse action explanation into the model, not after it
Choose a model architecture and reason code mapping that can explain a decline in plain terms from the start; retrofitting explainability onto an opaque model after launch is much harder.
- 3
Champion challenger every model change
Run a new model version against a control group before full rollout, and measure the change in both approval rate and downstream loss rate, not approval rate alone.
- 4
Monitor by origination cohort, not only in aggregate
Track delinquency and charge off by the month or week a loan was originated, so a deteriorating cohort is visible in weeks rather than showing up only in a lagging annual loss number.
Guardrails
- Credit policy and decision thresholds are version controlled and changed only through a documented approval process
- Every decline carries a specific, accurate adverse action reason a consumer can act on
- New model versions are tested against a control group before they set decisions for the whole book
- Approval rate and loss rate are reviewed together at every model or policy change, never one without the other
KPIs to instrument
- Approval rate and completed purchase rate, measured against a control group for any change
- Delinquency and charge off rate by origination cohort
- Adverse action accuracy, whether the stated decline reason matches the model's actual decision factors
- Time from a detected cohort issue to a policy or model change
Human in the loop
No individual checkout decision is reviewed by a person; the human role is model governance, not transaction review. A model risk function signs off on every new model version and policy change before rollout, reviews cohort level performance on a fixed schedule, and can force a rollback to a prior model version if a cohort deteriorates.
Common failure modes
- A model change is judged on approvals alone
- A new model approves more applicants and looks like a win before enough of that cohort has had time to default. Hold judgment until the cohort has matured, or use an early performance proxy validated against past cohorts.
- Explainability bolted on after the fact
- An accurate but opaque model gets a generic reason code added later that does not match what actually drove the decision, creating regulatory and consumer harm risk. Build the explanation into the model design.
- Thin file and no file segments drift unnoticed
- Performance on consumers with limited credit history is not tracked separately from the whole book, so a problem specific to that segment is masked by strong performance elsewhere. Monitor it as its own cohort.
What are the risks and rules?
EU AI Act
High risk
Annex III point 5(b): AI systems intended to evaluate the creditworthiness of natural persons or establish their credit score are high risk, and a real time BNPL underwriting model is squarely this use, since it is the system making the credit decision rather than a supporting tool. Providers need risk management, data governance, logging and human oversight; deployers must run a fundamental rights impact assessment (Article 27), and consumers have a right to an explanation of an individual decision (Article 86).
Controls to put in place
- Adverse action reason codes that match the model's actual decision factors
- Version controlled model and policy changes with a documented sign off before rollout
- Champion challenger testing against a control group for every model change
- Cohort level delinquency and charge off monitoring with a defined escalation path
Frequently asked questions
- Does a human review each buy now pay later approval?
- No. The decision is automated end to end at checkout speed. The human role is governing the model and policy: sign off on new model versions, monitor cohort performance and set the rules that turn a risk score into a decision, rather than reviewing individual transactions.
- What results have named BNPL providers disclosed?
- Affirm announced a transformer based underwriting model on 17 September 2026 and reported that its initial deployment produced 3.4% more completed purchases measured against a control group, approving eligible applicants, including some with no FICO scores, that its prior system would have declined. Sezzle uses a proprietary machine learning score, the Prophet Score, grouped into A to C bands, as the credit quality indicator for its receivables portfolio, last updated in October 2023.
- Is buy now pay later underwriting high risk under the EU AI Act?
- Yes, generally. Annex III point 5(b) makes AI that evaluates the creditworthiness of natural persons or establishes a credit score high risk, and a real time BNPL underwriting engine is the system making that decision, not a supporting tool, so the high risk obligations apply directly.
- How is this different from a bank's alternative data credit scoring?
- The underlying modeling techniques overlap, but BNPL underwriting decides in real time at checkout on a small, often fee free loan with thin margins, usually without an existing customer relationship, which puts more weight on speed, explainability at scale and rapid retraining than a bank's periodic credit decisioning process.
How to cite this page
Blits.ai AI Use Case Library, "AI underwriting and credit risk decisioning for buy now pay later", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/bnpl-underwriting-and-credit-risk. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 29 September 2026: First published