AI for Underwriting in Retail

A customer at a till taps through a store card application while a queue forms behind her, expecting an answer before the next item is scanned. Two floors up a credit controller decides whether to release a five figure order to a trade customer with three invoices past terms and a site starting Monday. Both are credit decisions. Neither sits with anyone whose job title says underwriter.

Two books, and only one of them looks like credit

Most retailers run two credit books at once. The consumer book is a store card, an instalment plan or a buy now pay later decision taken at a checkout or in an app: the applicant is an individual, the clock runs in seconds, and the retailer is often not the lender of record but an intermediary for a partner lender.

The trade book is credit accounts for wholesale, trade counter, foodservice and concession customers. There the applicant is a business with filed accounts, the output is a limit and payment terms rather than a single yes, expressed day to day as an order release queue.

They share a risk appetite and little else, so policy written for one gets applied to the other by accident.

Seconds at the checkout, and the basket that walks

In the consumer book the answer must land while the applicant is still standing there. A referral is not a delay, it is a decline with extra steps, because the basket is abandoned long before anyone reviews the case.

That rules out an agent in the scoring path and not much else. After the moment, the referral lane can be worked in minutes rather than days, with the case already assembled and a recommendation drafted, so a human accepts or rejects rather than researches. App journeys and higher value instalment plans tolerate a short wait, which is where a prepared referral converts.

A first application backed by four years of purchase history

Plenty of applicants are thin file or new to credit. The bureau has little to say, and the only substantial evidence belongs to the retailer: account tenure, order frequency, delivery address stability, payment instrument history, chargeback record, returns behaviour. An agent can turn that into a consistent dossier for every applicant, each item pulled from a named system and timestamped.

Fairness and lawful basis, not just accuracy

Behavioural data raises a question accuracy does not answer. Was it collected for a purpose that covers credit assessment, can you state the lawful basis, and does the field proxy for something you may not use? High returns rates correlate with disability in ways unrelated to willingness to repay. The field list is a policy document, not a modelling choice: write down which of your own items are admissible and which are excluded, and have the agent work only from the admissible set.

Affordability is a judgement you must be able to defend

Affordability and vulnerability obligations mean the decision must be defensible to someone who was not in the room: what the applicant declared, what the evidence showed, what policy required, and what happened when those disagreed. An agent is good at that assembly: declared position against evidence, conflicts highlighted, and the indicators of distress that policy says must stop an application rather than price it.

The reason you give and the reason you hold

A declined applicant is owed a reason that is both true and permitted to disclose. True rules out a template that does not describe what happened. Permitted rules out repeating a third party record you may not pass on, or describing a fraud control precisely enough to help someone testing it. An agent can draft the explanation and show the reviewer the policy clause it relies on.

Returns abuse is not credit risk and should not be priced as one

A checkout gives one answer, so fraud and credit get collapsed into a single no. Identity fraud, first party misuse, voucher abuse and serial returns are real losses, and none of them is an inability to repay.

Pricing them as credit risk creates three problems: an applicant who can afford the plan is declined on credit grounds, the reason given is untrue, and the outcome data that trains the next version of policy is mislabelled. Keep two decisions, two reason code sets and two owners. An agent can sort a referral queue by which question is actually open.

When you are the shopfront and someone else is the lender

Where the retailer introduces rather than lends, three questions decide how much work you own. Whose policy applies when yours is stricter than the lender's. Who answers the complaint, which in practice is always the shop. And how an evidence pack reaches the partner lender in the form that lender needs.

The last is where the time goes: every partner wants the same facts in a different arrangement. An agent can assemble the pack to the partner's format, keep your copy of what was sent, and track what came back. That record is your only defence in a complaint, and it is usually the thing nobody kept.

Every figure in the file should name its source

The trade file is a pile of documents that disagree: abridged accounts a year old, management accounts in a spreadsheet, bank statements as a PDF, your aged debt, two references chosen by the applicant.

An agent reads these and extracts the standing items rather than the headlines. From statements: rent, payroll, finance and lease payments, returned direct debits, days in excess, the balance pattern at month end. From accounts: the figures a controller would otherwise rekey, plus the gaps where a number is absent rather than zero. Every figure then names the document and page it came from, so a reviewer checks the extraction instead of repeating it, and a number nobody can source is treated as missing.

Reconciling trading names to the customer that owes you

Trade books accumulate structure. Several trading names, delivery accounts opened by different branches, a concession under its own banner, a franchisee whose invoices go to a holding company. Each has a limit. The parent has none, because nobody set one.

An agent can match the names and addresses on your accounts to registered entities, build the ownership tree, sum real exposure across ledgers including concession and consignment balances, and flag where the entity on the order is not the one you credit checked. It can also note which guarantees are on file, as against the ones everyone assumes.

Nobody asks for more credit early

A limit is a standing position, not a decision taken once. The business changes and the request arrives at the worst moment: the week before peak, from a sales manager, for a customer last reviewed two years ago.

Running the whole book continuously is the only version that works. An agent can check every account against review triggers each night, order pattern shifts, a returned direct debit, part payments appearing, a new entity in the group, a filing now late, and draft a recommendation with the reasoning stated and the evidence attached. The value is that the handful near the policy boundary are surfaced as such rather than found inside a seasonal uplift request.

A held order costs you a sale before it costs you a loss

The release queue is a commercial problem wearing credit clothing. A held order is a lorry not loaded and a sales conversation that has started badly. The question is whether it clears on evidence or on whoever shouts.

An agent can work it by proposing a disposition for each held order and showing its working: release inside the available limit, release against a payment sitting unapplied in the bank, part release to the value covered, hold pending one named document, or escalate with the reason stated. A controller accepts or overrides. Watch time to disposition, and how many overrides go the same way, which tells you whether the policy or the agent needs correcting.

The record has to outlive the policy and the approver

A credit decision has to be reconstructable months later, after the policy has changed and the person who approved it has left. So the record keeps the policy version in force, the inputs and documents used, what the agent produced, what the human decided, and where the two differed.

Overrides are the most useful data in the file: a steady pattern of controllers overriding in one direction is a policy problem, not an operator problem. Watch the spread of decline reasons too: a quiet shift there is usually the first sign something upstream changed shape.

Where AI fails here

An adverse credit decision about a person stays with an accountable human. There is no version of this where an agent declines an applicant on its own authority, or decides someone is unaffordable without a named person signing it off. The same applies to a limit reduction on a trade account, which can put a customer out of business and should never happen silently.

Affordability cases where the evidence contradicts itself need a human, not a tie break rule. So do vulnerability signals, which are often visible in the texture of a conversation rather than in a field.

First trade accounts with no filing history are badly suited to this: there is nothing to read, and a prepared file provides false comfort. The honest output is a note saying which evidence does not exist.

Outcome data arrives late. A limit set today shows its quality many months out, so a policy that looks like it is performing may simply be young. Measure on contested cases, not easy ones.

Finally, pressure to raise automated approval rates arrives during peak, when you have least capacity to review what changed. Treat threshold changes in trading season as policy decisions with named owners, not configuration.

Where to start

Pick one lane and baseline it before building anything. The trade order release queue is usually the best first candidate: the decisions are already written down, the evidence is already in your systems, and there is a clean number to improve. The consumer referral lane is the alternative if that is where conversion leaks.

Run it in shadow for a few weeks. The agent prepares the file and proposes a disposition, the human decides without seeing the proposal, and you compare. Agreement on easy cases tells you little. Agreement on the cases that went to a second opinion tells you whether the preparation is trustworthy, and disagreement there is what you will learn most from.

Adopt AI Partners builds this kind of prepared file and queue agent inside the ledger, bureau connection and order system a credit team already uses, so the decision record lands where auditors already look. If a release queue or a referral lane is slower than it should be, describe the situation to hello@adoptaipartners.com and we will tell you whether it is worth handing over.