AI for Underwriting in Consumer Goods

A credit application arrives as a scan attached to a sales email. A wholesaler trading under a name nobody on the credit desk recognises wants terms on an opening order of mixed ambient lines. The form carries a company name, a trading address, a VAT number, two trade references and a signature. The box for the registered number is blank. The email above it says the buyer wants delivery next week and the area manager has promised launch support.

Nothing in that file answers the question you are actually being asked, which is how much unsecured credit your company will carry against a business you cannot see, in a channel where margins are thin and a failure takes the stock with it.

Trade credit underwriting in consumer goods is not pricing a risk. It is sizing an exposure and then watching it, across hundreds of accounts, on a ledger that changes every day.

Small accounts arrive with abridged accounts and two trade references

Independent retailers, single site foodservice operators and family wholesalers do not come with rated financials. You get abridged accounts with no profit and loss, filings for a period in which the company barely traded, or nothing because the entity was formed last spring. The agency score sits on the same thin data plus a payment file your own ledger would improve on.

What takes the time is assembly. The registered entity and whether it is the entity on the order, the directors and their other appointments, late filings, charges registered, judgments, prior insolvencies among the same directors, and the group the applicant sits inside. An agent can do that assembly and hand over one sheet with every source named and dated, so the underwriter reads evidence instead of hunting for it.

Spreading filed accounts that describe a year already closed

Where real accounts exist, they cover a period that ended long before the application landed. They are still worth spreading: trade creditor days against your proposed terms, stock turn, net current assets, directors' loans moving in or out, the auditor's wording, whether the accounting reference date was changed to buy time. Extraction and ratio work is mechanical. Whether thin net assets matter in a wholesaler whose directors traded through two downturns is not.

Trade references say more about who was asked than about who was paid

Applicants choose their referees, so the suppliers named are usually the ones who get paid first. That does not make references worthless, but it changes what you look for in them.

An agent can draft the requests, chase the unanswered ones, read replies in whatever form they arrive, and normalise them into the same fields: how long the account has been open, highest credit granted, terms actually operated, oldest balance outstanding, whether the account has ever been stopped. It should also say what the set omits. No referee in your category, nobody supplying at the volume requested, two referees sharing a registered office with the applicant.

Five trading names, one parent, and an exposure nobody has added up

Symbol group members, franchise operators, foodservice groups buying through several sites and wholesalers with separate depot accounts create the same problem. Each account was opened at a different time by a different salesperson, each has its own limit, and all of them roll up to one controlling company. The aggregate is the exposure you are really carrying.

An agent can maintain that mapping continuously rather than at audit time, matching on registered number, registered office, common directors, delivery addresses and the bank details on remittances, then report the rolled up balance against a group limit. The inverse error matters too: accounts wrongly linked, an application refused because a namesake failed elsewhere.

A limit set before the season is the wrong limit during it

A limit sized in February against a February order pattern is the wrong number in October when the customer is building Christmas stock. The same happens ahead of a barbecue season, before Easter, and around any on pack promotion that triples a line's volume for a few weeks. Sales finds out when the order hits credit hold, and the request arrives as an urgent uplift.

This is foreseeable, which makes it good agent work. Compare each account's order pattern against the equivalent period last year, project the peak balance against the limit and the insured cover, and list the accounts that will breach before they do. The uplift then gets decided from a prepared file rather than over the phone with a lorry waiting.

The ledger changes shape before a payment is missed

Most accounts that go bad pay you for months first. What changes is the texture. Payment moves from early to the last day of terms, then just past it. Invoices get part paid rather than cleared. The settlement discount is taken and the payment arrives outside the window. A direct debit is represented. A cheque is returned unpaid. Someone starts paying by card for exactly the amount needed to release the next delivery.

Each signal is weak alone. Together, across a ledger with hundreds of live accounts, they are the earliest warning you have, and nobody has time to read them account by account. An agent can read the ledger nightly and return a short ranked watch list with the evidence attached, which is a different artefact from an aged debt report sorted by days overdue.

Deductions and disputes are where an inability to pay hides

A customer who cannot pay often argues instead of defaulting. Short delivery claims appear, then damage claims, then price queries on invoices priced the way they always were, then promotional deductions against agreements nobody signed. Deduction volume rising on an account whose service level has not changed is a credit signal before it is a logistics one.

An agent can reconcile each claim against the proof: the delivery note and the signature on it, the price list effective that day, the signed promotional agreement and its claim window. That separates a genuine dispute from a balance being held back, which your overdue report cannot do.

What the order book says about an account in trouble

Order behaviour carries information the finance data does not. Order frequency rising while order value falls. A switch towards lower value lines. Orders arriving just below the value that triggers a release check. Delivery requested to an address that is not the trading address. And the pattern everyone recognises afterwards, the unusually large order from an account that has never bought at that scale, placed shortly before the business fails.

An agent watching order entry alongside the ledger can raise these as they happen. That last order looks like good news to sales, and only looks like a warning to someone comparing it against the account's own history.

The cover schedule is not your limit list

If you buy credit insurance, there is a cover schedule somewhere that is already out of date. Limits the insurer granted, limits it quietly reduced, cover withdrawn by endorsement, accounts trading under your discretionary limit where no cover was requested, conditions your collections process is supposed to satisfy.

Reconciling that schedule against live balances is clerical, important and almost never done weekly. An agent can match insured limits to internal limits and current exposure, report the genuinely uninsured amount, list accounts where granted terms exceed what the policy allows, and flag cover reduced months ago against an internal limit never adjusted to match.

The release queue, and the call sales wants within the hour

The order release queue and the credit hold list are where underwriting actually bites. Orders stopped for over limit, for overdue balances, for being a new account with no limit yet. Most are routine. All are urgent to somebody.

An agent can prepare each held order before the desk opens: balance, overdue amount and its age, the limit, remaining headroom, what this order adds, whether a payment has cleared since the hold went on, whether the overdue item is a deduction already under review, whether insured cover exists, and a recommendation with the policy clause behind it. Releasing stays with a person, but arriving to a prepared queue changes what that person does with the morning.

Reviewing hundreds of limits without reading hundreds of files

Policy says every limit is reviewed annually. In practice reviews happen when something goes wrong, and the quiet accounts carry limits set by somebody who left.

A continuous sweep suits an agent well. Refresh agency data, read new filings, recompute payment behaviour from your ledger, check the group mapping, compare the limit against exposure and cover, then sort every account into no change, propose an increase, propose a reduction, or needs a human. The credit manager spends the month on the accounts that moved.

Where AI fails here

Appetite is not an analytical output. Granting terms against a thin file is a commercial decision about how much of a channel you want and what you are willing to lose to get it. An agent can give you the exposure, the comparable accounts and the policy position. It cannot tell you whether this customer is worth the risk.

Withdrawing terms from a customer sales fought to win stays a human conversation. So does stopping an account, agreeing a payment plan, and deciding a balance is now a collections problem rather than a credit one.

Fraud built to look ordinary will defeat the model. An account that pays perfectly for a year to build its limit, then places one large order and disappears, looks exactly like a good customer until the last week. Nothing in the data distinguishes patience from intent.

The agent cannot repair your customer master. Duplicate accounts, trading names held as separate entities and missing registered numbers will produce confident group aggregations that are wrong. It shows you the inconsistencies, but fixing them is your work and usually the real prerequisite. Nor can it make an insurer raise a cover limit, make a customer file accounts, or make a referee reply.

Where to start

Take the credit hold list for one channel and have an agent prepare it, not decide it. Every held order arrives with balance, ageing, headroom, cleared payments, open deductions, cover position and a recommendation with its reasoning. Release authority does not move. Within a few weeks you will know what share of holds were genuinely borderline, and that figure says more about your limit quality than about your queue.

The second build is usually the behaviour watch list over the ledger, because it finds exposure you were carrying without knowing. Group aggregation follows, once the customer master is clean enough to trust.

Adopt AI Partners builds these agents inside your existing ledger, order entry and credit systems, so a prepared file lands on the account rather than in another tool. If you want to talk it through, describe how a new terms application reaches your desk today to hello@adoptaipartners.com.