TL;DR
- What it is: AI product recommendations score which financial product a specific customer is likely to want and qualify for next, based on their own behavior, not a generic segment.
- The distinction that matters: a product recommendation, what to offer, is a different decision than next-best-action, whether, when, and how to act on it.
- Why it has to explain itself: a recommendation that touches a credit decision has to say why, alongside performing well.
- The harder problem: a new-to-bank customer with no transaction history breaks a model built only on behavioral signals.
- The economics: banks spend more acquiring new customers than activating existing ones, and the typical customer already holds less than half their financial relationships with one bank.
- Honest caveat: recommending the right product doesn’t sell it. The offer still has to reach the customer through a channel and moment that fits.
Most conversations about AI in banking marketing use “next best offer,” “next best action,” and “product recommendation” as if they were interchangeable. They aren’t. A recommendation engine answers one question: which product should this customer see next. A next-best-action system answers a bigger one: whether to act at all, through which channel, at what moment, and with what, a product offer or something else entirely, a retention call, a fraud alert, a financial-wellness nudge.
AI product recommendations in banking are the narrower, more concrete of the two: a model scores a specific customer against a specific product using their own transaction and product-holding history, not a population average. Get that scoring right, and the next-best-action layer built on top of it has something worth acting on. Skip the distinction entirely, and a bank ends up sending the same offer to everyone regardless of channel or timing.
The build-validate-deploy sequence behind any of iTuring’s predictive models applies here too. What’s specific to recommendations is what happens after the model scores well: explaining the result, handling a customer with no history yet, and proving the economics are worth the build.
Next-best-product and next-best-action are not the same decision
Next-best-product is a recommendation: which product a specific customer is likely to want and qualify for. Next-best-action is a broader decision: whether to act on that customer at all, and if so, through which channel, at what time, with what message, a product offer, a service intervention, or a fraud alert.
One of the clearer industry definitions treats next-best-action as a decision made for a segment of one, in real time, from live behavior, and covers service interventions and fraud alerts, well beyond cross-sell messages. Plenty of vendor content, including from major analytics and consulting firms, uses next-best-action and next-best-offer interchangeably anyway. That’s a modeling mistake as much as a semantic one. A fraud alert and a credit-card offer have no business running through the same decision logic, timing rules, or approval chain.
This piece stays scoped to the narrower, more concrete layer underneath all of it: the product recommendation itself.
Why explaining a recommendation matters as much as making it
A retailer recommending a blender never has to justify it. A bank’s recommendation, where it touches a credit decision, a pre-approved loan offer, a credit limit increase, sits closer to a regulated outcome, and regulators in several markets have already made clear that a vague reason won’t satisfy that obligation.
In the US, the Consumer Financial Protection Bureau’s 2023 guidance on adverse action notices spelled this out directly: if an algorithm factors in something like an applicant’s profession, a generic reason such as “insufficient projected income” doesn’t satisfy the disclosure requirement. The reason has to be specific enough that the applicant understands what actually drove the decision. That obligation bites hardest where a recommendation feeds a credit decision. A pure savings or investment cross-sell sits in a lighter-touch zone, though it still carries ordinary model-risk expectations around explainability and audit trail.
This is exactly why explainability has to be built into the model, not bolted on afterward with a separate interpretation tool applied after the prediction is already made. iTuring’s platform carries explainability at both the model and the individual prediction level, with the same lineage and maker-checker approval every model on the platform goes through, so the evidence exists by default rather than being reconstructed after a regulator or a customer asks for it.
The cold-start problem is worse in banking than in retail
A new e-commerce shopper with no purchase history just sees popular items, low stakes either way. A new-to-bank or thin-file customer is the exact segment where a wrong or absent recommendation matters most, and where falling back on demographic data alone risks proxying for creditworthiness in ways a fair-lending review would catch.
One practical workaround, described in a well-regarded critique of market-basket-style banking recommenders, sequences products by a needs hierarchy instead of a single similarity score: a deposit account first, then a debit card, then products that build engagement and financial stability, and only then loan products matched to life stage. That ordering gives a thin-file customer a sensible next step even before there’s enough transaction history for a behavioral model to work with.
The same logic applies to the model itself. A recommendation engine that leans harder on account and onboarding data when transaction history is thin, and shifts weight to behavioral signals as that history builds, handles a new customer more honestly than one that either ignores them or guesses from demographics alone.
The economics that make cross-sell worth building for
Bank marketers spend roughly 34% of their budget on acquiring new customers and just 23% on retaining existing ones, according to a 2025 ABA Banking Journal survey, an imbalance the same research ties to a simple fact: the typical customer holds less than half their deposits or loans with any single bank.
The same analysis argues that marketing to existing customers returns something like 10 times the ROI of new-customer marketing, with response rates 5 to 10 times higher. That’s the report’s own estimate rather than an independently verified industry figure, but the underlying logic holds regardless of the exact multiple: a customer a bank has already onboarded, already verified, and already holds transaction data on is a fundamentally cheaper problem to solve than a new one.
None of this works without the behavioral data underneath it. iTuring’s Data Accelerator draws on transaction, product-holding, and servicing signals with the lineage back to source that a recommendation feeding into a credit-adjacent offer needs to hold up to a model risk review.
If your recommendation engine can’t tell a data science team why it picked what it picked, it isn’t ready for a credit-adjacent offer yet. Book a working session with our data science team to see what an explainable recommendation layer would catch that yours doesn’t.
Sources
- Consumer Financial Protection Bureau, Circular 2023-03, “Adverse Action Notification Requirements and the Proper Use of the CFPB’s Sample Forms Provided in Regulation B,” September 19, 2023.
- CFS Insight, “Next Best Product (NBP): The Right Recommendation?” 2026.
- ABA Banking Journal (Mark Gibson, Capital Performance Group), “Effective cross-selling: The key to meeting deposit and loan growth goals,” March 24, 2025.
- iTuring platform pages: Data Accelerator, Model Gov, confirmed by direct fetch 2026-09-21.
