Cross-sell propensity at account opening

Cross-sell propensity at account opening is a ranking of one optional next product after the primary account exists. Finish KYC, screening, and booking for the product the customer applied for. Then, and only then, decide whether a single add-on is worth showing. The customer can refuse. The account stays open either way.

Do not delay identity, proof of address, or screening so a bundle can be assembled. Those checks belong to the product that is being opened. Holding them for a travel card, a fee package, or an overdraft turns onboarding into a sales queue. The applicant asked for a working account. Quality here means that account is real before anyone talks about the next product.

Bundling before the current account exists is the failure that looks efficient on a branch screen and fails in the file. The banker presents a current account plus a card plus protection as one decision. If the applicant declines the extras, the opening can stall, or the decline can be written against the wrong product. A later complaint then reads as if the bank refused the account. Separate the decisions: book the primary product on its own clock, then make an offer that can be refused without undoing the booking.

Propensity-to-open scoring is a different control. It ranks who is likely to complete an application. Do not reuse that rank at the desk as a next-product list. A person who was likely to open is not, by that fact, a candidate for every ancillary product in the catalog.

Build the rank from data you already hold with consent for this opening or for the existing relationship: customer segment, product holdings, and cash-flow signals collected to open the primary account. Holdings tell you what is already on the books, including products opened in another channel. Cash-flow tells you whether a savings sweep, a packaged debit card, or a small buffer product is even plausible. Segment keeps student, mass, and affluent propositions from mixing.

Use only fields the applicant agreed you could use for servicing and product decisions, plus holdings the bank already records. Do not enrich the rank with campaign clicks, purchased audiences, or inferred lifestyle labels that were never part of the application. If a signal was collected for marketing, keep it in marketing.

Do not rank a loan from a marketing score. A campaign propensity is not capacity, affordability, or willingness to take credit. If the next product is credit, the customer is at a point of credit intent only when they ask for it, and eligibility sits in credit policy. Real-time pre-approval at point of intent is the control for that moment. Thin-file credit limit calibration is the control for a limit if they accept a credit add-on. The opening rank only says which optional product is worth showing. It is not an underwrite of the add-on.

Lookalike audience generation belongs upstream in acquisition. Do not import those audiences as the ranked list at opening. A lookalike can tell you who to invite. It cannot tell you which product this applicant should be offered after the account is live.

CRM and core platforms, including Salesforce and Temenos, already hold relationship, holdings, and product-catalog state. Treat that class as the system of record for what the customer holds and what they are eligible for. Do not maintain a second product list in a worksheet the branch cannot refuse against, and do not let a channel overlay invent holdings the core does not show.

If the applicant declined a product in this session, on this application, or in a cooling-off window you already record, remove it from the rank. Re-offering a refused overdraft because the model still likes it is a quality failure. The customer already answered. Showing it again trains them to treat every screen as a pitch.

Apply the same rule across channels. A digital decline during the application must be visible to the banker who completes the booking. A branch refusal must be visible if the customer finishes on the phone or in the app later the same day. The rank should read the refusal file before it reads the propensity score.

Existing holdings are a refusal of a different kind. Do not offer a product the customer already has, even if the holding sits in another brand or channel of the same group and the model has never seen it. Eligibility checks exist to catch that. Run them before the offer is visible, not after the banker has started the script.

After the primary product is accepted, rank the remaining add-ons, then show one. Eligibility must already be true for that product: age, residency, existing holding, product-specific exclusions, and any rule that would stop origination if they said yes. If the top-ranked product would fail those rules, skip it and take the next one that would pass. Do not display a product and then discover, after a verbal yes, that the catalog would reject it.

A ranked list of five on a branch screen trains the banker to pitch the stack. One optional product, already checked, is the control. The script is short: the account you applied for is open; there is one optional product you qualify for; you can say no. Digital should use the same facts, not a carousel of add-ons that only exist because the model produced a top-n list.

Walk one illustrative path. A salaried applicant completes KYC and books a current account. Consented application data already shows a savings holding from a prior relationship and regular salary credits used to evidence the opening. The rank places a packaged debit card above a personal loan and above an overdraft the applicant declined earlier in the same session. The loan is not shown: a marketing score had favored it, but there is no consented credit application and no pre-approval. The declined overdraft is not shown. The banker names the packaged card, confirms eligibility against the catalog, and records a yes or a no. The current account is already live. Nothing in that sequence waits on a bundle, and the score is not used to book the card.

That is the operating sequence to keep: wait until the primary product is accepted, rank add-ons from consented holdings and cash-flow you already have, and show one optional offer with eligibility already checked.

A no is an outcome. Write it to the relationship so the next channel does not repeat the same product. Include enough context that a later model does not treat silence as missing data: product, channel, and that the customer refused after the primary account was open.

A yes is a handoff, not an approval. The add-on still follows its own origination path, disclosures, and, if it is credit, underwriting. If you treat the propensity output as an underwrite of the add-on, you will book products the credit or product team never agreed to, or you will skip disclosures because the opening screen already felt like a sale. Keep the score in the offer layer. Let origination own the contract.

Staff incentives should follow the same split. Do not pay the opening as a package when the add-on can still be refused. If the banker is measured on bundle take-up at the same moment as account booking, the process will collapse back into bundling before the current account exists. Measure a completed primary opening first. Count an accepted add-on only after origination has taken it, and count a recorded refusal as a completed offer, not as a miss.

When the file is clean, a product or branch-sales lead can audit a day of openings and see three facts on each one: the primary account was booked without waiting on a bundle, at most one optional product was shown after eligibility was checked, and every refusal is sitting on the relationship so it cannot be pitched again.