Personalise the offer, then prove the margin

A cross-sell campaign can change wording and timing while leaving the product-selection rule untouched. Before generating additional variants, review whether the offer matches an eligible customer’s needs and existing products. The operating decision concerns who receives what, not only how the message reads.

More message variants may improve presentation without improving offer selection. Assess the eligibility policy and the economics after purchase. This identifies where generation could help and where the underlying data or product choice needs a different remedy.

Choose a specific product gap, define eligible customers and preserve a randomised holdout under the existing process. Personalisation may improve matching, but the commercial result needs to survive contact costs, opt-outs, service effort, returns and the product’s actual margin.

A customer record can become an offer decision

The buyer for this decision is a chief customer officer, chief marketing officer, head of growth or retail banking director. Their workflow starts with an eligible customer list, chooses a product and channel, creates the message, delivers it and waits for a conversion. A first-party personalisation pipeline can add a customer representation built from transactions, behaviour and existing products, then use that representation to choose a relevant offer.

The useful boundary is a decision record that says why this person is eligible for this product now, which source signals support the choice and what the customer saw. The model can rank an offer or draft language. The growth owner still decides the allowed products, exclusions, contact policy and measure of value.

The freshest evidence comes from Casero and colleagues at Columbia Business School. Their working paper describes two field experiments with approximately 330,000 customers of a global financial institution. Customers received either the bank’s existing cross-sell offer or an offer generated by a pipeline using first-party behavioural and transactional data. Success was measured with actual conversion, rather than a model score. Casero and colleagues, Customer Digital Twins in the Field

The paper reports a 6 per cent conversion uplift and an 11 per cent prospective three-year gross-margin uplift for credit-card offers. For payroll-portability offers it reports a 25 per cent conversion uplift and a 30 per cent prospective three-year gross-margin uplift. These are prospective margin calculations from one financial institution, not realised results that a new programme can promise. The Columbia working paper

1. 01 Select Define the customer, product gap, eligibility rule and exclusions from first-party records.
2. 02 Personalise Use the permitted history to rank an offer and draft a message with its supporting signals.
3. 03 Deliver Send one offer through the chosen channel and preserve what the customer received.
4. 04 Observe Record conversion, margin, opt-outs, complaints and service cost against a holdout.
5. 05 Decide Expand, narrow or stop the offer policy when the incremental economics are clear.

Relevance is an economic mechanism, not a tone of voice

Personalisation earns its place when it reduces a customer’s search or choice cost. A relevant product can arrive when the account history shows a plausible need, with an explanation that uses facts the customer recognises. That can raise purchase incidence and reduce the chance that the first purchase is a poor match.

Evidence from a separate field experiment helps isolate the mechanism. A Wharton working paper studied personalised product rankings on a large online retail platform. Among 635,267 consumers, personalisation increased conversion probability by 1.4 per cent, revenue by 2.1 per cent and profit by 1.5 per cent. Repeat purchases increased two to 3.9 per cent at later observation points, and customers who bought a personalised product were 10 per cent less likely to return it. The experiment changed ranking based on browsing history, so it is evidence about matching and search friction, not about an LLM or a financial cross-sell offer. Korganbekova and collaborators, Personalisation and Consumer Welfare

The studies provide evidence for testing matching policies in their respective settings. They do not establish a general margin benefit. A local comparison must determine whether the offer creates incremental business and whether later costs or retention effects offset it.

Start with the product gap you can explain

Before choosing a model, choose one product gap and one customer population. For each offer, write the eligibility rule in ordinary language. A customer without the product, with a relevant transaction pattern and with consent for the channel might qualify. A customer with a recent complaint, a pending service case or a conflicting exclusion should not be placed in the same queue by default.

Build the baseline from the current campaign. Record the offer, price, channel, send frequency, conversion definition, gross-margin calculation and service cost. Preserve the source signals used for the personalisation decision. If a reviewer cannot explain why a customer entered the treatment group, the system has created a score rather than an accountable offer policy.

The holdout should last long enough to observe the action the business cares about. A credit-card application, a funded account or a product renewal is a different outcome from a click. Keep the randomisation at the customer level when offers can spill across channels, and record every contact so a second campaign cannot quietly contaminate the comparison.

Keep the customer model inside a control boundary

First-party data does not make every use legitimate or useful. A transaction history can reveal a need, a constraint or a sensitive circumstance. Decide which attributes may influence offer selection, which may appear in customer copy and which must remain out of both. Give the customer a way to decline the channel, and give an operator a way to remove an offer when the underlying data is wrong.

The pipeline should preserve a reviewable record: the eligible customer set, product gap, source fields, model output, message, delivery event and later outcome. Keep original data separate from generated summaries. An explanation is not proof that the source signal was present or that the offer was suitable.

Ask a supplier to demonstrate stale and contradictory records. Test an account that closed yesterday, a product that is already active in another system and a customer who opted out after the list was built. The correct result is a held or withdrawn offer with a reason, not a plausible paragraph sent on schedule.

What the Inference Institute can help decide

The useful engagement starts with the offer workflow and the margin constraint. We map the data that creates eligibility, separate the model’s ranking from the business policy, design the holdout and write the evidence needed to review the result.

The output is a pilot specification a growth or customer leader can fund. It names the first product gap, population, channel, treatment boundary, observation window, economic measures and stopping rule. It also makes clear which later change would justify widening the policy and which would require better data or a different offer.

What this does not tell you

The Columbia study is a September 2026 working paper. Its experiments concern one global financial institution, and its gross-margin figures are prospective calculations rather than realised profit. The Wharton study concerns product rankings on an online retail platform and uses browsing history. Neither study establishes that an LLM, a digital-twin design or a particular vendor will produce the same outcome in another organisation.

Neither paper removes the need for a local control period. Personalisation can increase contact pressure, expose weak product data or optimise a metric that does not survive cancellation and service costs. A positive conversion result without an incremental margin calculation is an incomplete business case.

Before commissioning more generated copy, identify a supported product gap and the holdout result that would justify expansion or stopping. Preserve eligibility evidence and later economics so the owner can evaluate the offer policy rather than infer its value from engagement.