No. A buy now pay later approval is a credit decision. On the lender's side, an approved credit model and policy make it, and the AI agent gathers data, explains and escalates. On the shopper's side, an agent buying on a customer's behalf must never take on credit for them without the customer's own clear agreement.
Why is a buy now pay later approval a credit decision?
Buy now pay later feels like a payment option, but it is lending. The customer receives goods now and owes money later. If they cannot pay, they face late fees, collection activity and, in many markets, an effect on their credit record. Several small plans taken out together can add up to more debt than the customer realises.
That means a BNPL approval carries the duties of a credit decision. The EU AI Act lists creditworthiness assessment and credit scoring of natural persons as high-risk. GDPR Article 22 and its UK equivalent restrict decisions based solely on automated processing with legal or similarly significant effects. Where those rules apply, the decision needs the safeguards they require, whoever or whatever makes it.
Who should make the decision on the lender's side?
BNPL decisions are made at the checkout in seconds, so some automation is unavoidable. The question is what kind. A credit model that has been built, validated and approved under the lender's model governance, with a written policy around it, is the right decision-maker for routine approvals. A general-purpose agent reasoning its way to a yes or no is not.
So on the lender's side, the agent's role sits around the decision:
| Step | The agent | Stays with the model or a person |
|---|---|---|
| Gathering data | Collects the application, checks identity and pulls the data the model needs | Nothing routine |
| Deciding | Passes the case to the approved model and records the result | The model, under policy; a person for referrals |
| Referrals | Prepares the case with what triggered referral | A credit officer decides |
| Declines | Drafts the reasons from the model's recorded factors | Reconsideration by a person |
| Explaining | Explains the plan, fees and payment dates to the customer | Disputes |
| Limit changes | Prepares the case | The model or a person, under policy |
What about an agent shopping for a customer?
Agents that browse and buy on a customer's behalf raise the other side of the question. If the agent finds a BNPL option at checkout, should it accept?
No. Choosing credit is a decision the customer must make. The agent can present the option, the cost and the payment schedule, and ask the customer. It should never select a credit option to complete a purchase because it falls within a spending budget. Paying a customer's bills raises a similar question, covered in whether an AI agent can pay a customer's bills.
A worked example
Take a BNPL lender working with online retailers. A customer chooses to pay in four instalments for a purchase of 400.
The agent checks the customer's identity and existing plans with the lender. The customer already has three active plans. The agent sends the case to the approved credit model, which refers it because total outstanding balances would exceed the policy threshold for the customer's profile. The agent tells the customer their application needs a short review and shows the time it may take. A credit officer reviews it, sees two plans are nearly paid off, and approves.
Two months later, the customer misses an instalment and writes that they lost their job. The agent records the promise they make to pay part of the amount, as described in handling promises to pay with an AI agent, and passes the case to the financial support team.
What controls make it safe?
- The model decides, the agent supports. The agent's role scope includes running the approved model, never substituting its own decision.
- Policy as written instructions. The lender's credit policy becomes the agent's instructions, so a change to policy is a change to the agent. That is the argument in your procedures are now the agent's instructions.
- Reasons recorded at decision time. Each decline carries the factors the model recorded, so any explanation is grounded in the real decision.
- Customer consent for agent shopping. An agent acting for a shopper requires an explicit confirmation before any credit is taken.
- A full record. Each application, model result, referral and explanation is filed.
The wider limits are set out in AI agents in banking: what they may and may not do.
Questions people ask
Can the agent raise a customer's BNPL limit?
Raising a limit extends new credit and is a credit decision. The approved model and policy decide, with a person for referrals.
Should customers know an agent is involved?
Yes. People should be told when they are interacting with an AI system, and a customer taking credit has a strong interest in knowing how the decision was made.
Can the agent waive a late fee?
Within a set policy and limit, it can. Waivers outside policy are drafted for a person, who can also consider whether the customer needs more help than a waiver.