Next-best-action: personalisation beyond the storefront

Next best action AI answers a different question from a product recommender. A recommender asks "which item?"; a next-best-action model asks "which intervention, if any?" The candidates are not products but things the business can do: send an offer, prompt a feature, schedule a call, waive a fee, offer a top-up, or leave the customer alone. It is the form personalisation takes in banking, insurance, telecoms, subscription software and any business whose relationship with a customer outlasts a single purchase. This article explains how NBA models are built, why the action catalogue and the feedback loop matter more than the algorithm, and how to start without boiling the ocean.

What a next-best-action model is

At its core it is a scoring system. For each customer and each eligible action, the model estimates the expected value of taking that action now, measured against an objective the business has chosen: retention, lifetime value, product adoption or a blend. The action with the highest expected value wins, subject to constraints such as contact frequency, regulatory eligibility and channel capacity. "Do nothing" is always a candidate and often the winner. The estimate can come from several sources: propensity models per action, uplift models that predict the change in outcome caused by the action, and bandits that learn from the outcomes of recent decisions.

NBA models compared with recommenders and rules

The action catalogue: the part everyone underestimates

The model can only choose among actions someone has defined. A good catalogue lists each action with its cost, its channel, its eligibility rules, its cooling-off period and the outcome it is supposed to move. A bank might list a credit-limit review, a fee waiver, a savings product nudge, a fraud-check call and a dormant-account reminder. A SaaS company might list an onboarding prompt, a feature tip, a plan-upgrade offer, a customer-success call and a renewal reminder. Building this catalogue with the teams who execute the actions is most of the work in the first month, and it is where the constraints that keep the system safe are captured.

Churn intervention AI: the save as an action

The most common first use of NBA is retention. A churn prediction model produces a risk score; on its own that is a list, not a decision. NBA turns it into one by choosing, for each at-risk customer, which save to attempt: a discount, a call, a feature walkthrough, or none because the customer is unlikely to respond to anything. Uplift modelling matters especially here. A discount sent to someone who was going to stay costs margin, and a call to someone who was going to leave regardless costs a customer-success hour. The uplift model directs effort to the persuadable middle, which is where retention budgets earn their return.

Next best offer in banking and insurance

Financial services was the first home of NBA because it has many products, long relationships and strict rules about who can be offered what. The model sits on top of eligibility: it never proposes a product the customer does not qualify for, never exceeds contact limits, and records the reason for every decision so a compliance reviewer can see it. Regulators expect fair treatment and explainable offers; the Reserve Bank of India's guidance on digital lending and customer protection is one reference point for lenders. In practice the model's job is to pick the best among the permitted actions, and the permitted set is defined by people. Where the models run on customer data that cannot leave the bank, we deploy them as private, self-hosted AI on the bank's infrastructure.

Beyond retail and banking

• Telecoms: top-up prompts, plan changes and network-issue apologies chosen by likely response
• Healthcare: appointment reminders, follow-up nudges and care-programme enrolment, within clinical rules
• Education: intervention for a learner falling behind, chosen from tutoring, reminders or a lighter path
• Logistics and marketplaces: incentives to drivers or sellers to cover a shortfall, priced by predicted response
• B2B SaaS: the onboarding prompt, upgrade offer or success call most likely to move a given account

The feedback loop is the product

An NBA system without a closed loop stops learning on the day it launches. Every decision must be logged with the customer, the action, the score and the alternatives considered; every outcome must be joined back to that decision within a defined window. That requires the campaign tools, the CRM and the call centre to report what was executed and what happened, which is an integration task, not a modelling one. It is also why a small random share of decisions should be taken at random rather than by the model: without exploration the model cannot learn about actions it never tries, and without a random control there is no proof of lift, as set out in why you must run a controlled test.

Guardrails: policy-gated actions

Some actions carry cost or risk: a fee waiver, a large discount, a regulated product offer, an outbound call. We treat these as policy-gated actions, the same stance we take with every AI agent. The model proposes; a policy layer checks eligibility, budget, contact frequency and any human-approval rule; only then does the action execute. Actions that fail the gate are logged as such, and the gate's rules are owned by the business, not the model team. Run the system in shadow mode first, recording what it would have done alongside what the current process did, so the teams can review its judgement before it acts.

A worked example

A subscription software company had a churn score and a customer-success team that worked through the list from the top. Everyone on the list got the same treatment: a call. We built an action catalogue with the success and marketing teams: a call, an in-app walkthrough for an underused feature, a short discount at renewal, a usage-report email and no action. Historical campaigns had some accidental randomisation, enough to train an initial uplift model per action, and we added a small random arm going forward. The system ran in shadow mode for a month while the team compared its proposals with their own. When it went live, calls were reserved for accounts where a call was predicted to change the outcome, walkthroughs went to accounts whose churn signal was low feature adoption, and a meaningful share of at-risk accounts received nothing because nothing was predicted to help. The success team covered more accounts with the same hours and the retention result held up against the random control. The delivery pattern follows the one used for the in-app copilot for a B2B SaaS, where the in-app surface carried the walkthroughs.

Team and timeline

An NBA build needs a data scientist for the uplift and propensity models, a data engineer for the decision log and feedback joins, an integration engineer for the CRM, campaign and call-centre connections, and business owners for the action catalogue and policy rules. A first version covering one objective and five to eight actions takes eight to twelve weeks including shadow mode. It is delivered under personalisation engines, from $21,000 / ₹13.6L, with the underlying churn or propensity models under AI/ML development at $17,500 / ₹11.2L where they do not exist. A Sprint Zero at $3,250 / ₹2,00,000 is the fastest way to build the action catalogue and confirm the data supports uplift modelling before committing. Care Plans on the pricing page cover retraining and monitoring.

Before you start: a checklist

• Choose one objective for the first version and write down how it is measured
• Build the action catalogue with cost, channel, eligibility and cooling-off period per action
• Check whether past campaigns had any randomisation to train uplift models from
• Confirm CRM, campaign and call-centre tools can report executed actions and outcomes
• Define the policy gate for costly or regulated actions and who owns its rules
• Reserve a random decision arm for exploration and proof of lift
• Plan a shadow-mode period where the teams review the model's proposals
• Agree the decision log format with compliance where regulated products are involved

Glossary

• Next best action (NBA): the intervention with the highest expected value for a customer now, including no action
• Propensity model: predicts how likely a customer is to respond to an action
• Uplift model: predicts how much an action changes the outcome compared with not acting
• Bandit: an algorithm that balances trying actions to learn with choosing the best known action
• Action catalogue: the defined set of actions with their costs, rules and intended outcomes
• Policy gate: the rule layer that checks an action before it executes
• Shadow mode: running the model and logging its decisions without acting on them

Related reading

Read churn prediction: building a model your retention team will use for the scoring side and personalised messaging for the delivery side; for banking, the fintech industry page describes how we handle data residency and audit. Pricing for every service is on the pricing page.

Define the actions, close the loop, gate the risky ones, and let the model choose; that is next-best-action done in a way a compliance officer and a finance director can both live with.