AI-driven retention that tripled repeat visits on a loyalty platform

Approach

We started by auditing the signal Ferry already had — redemption history, visit cadence, offer open rates, merchant category — and mapped it against the moments where retention actually breaks: the second visit, the lapsed-month mark, and the post-promo drop-off. Rather than bolt on a generic recommender, we framed the problem as a lifecycle agent: a service that decides, per customer per day, whether to stay quiet, surface an offer, or escalate to the merchant dashboard for a manual nudge. That framing kept the model’s job narrow and measurable.

Solution

We shipped a retention engine with three moving parts. A feature pipeline turns raw events into churn-risk and affinity scores, refreshed hourly. A next-best-offer agent — an LLM wrapped around structured tool calls into Ferry’s catalogue — picks the right deal from the merchant’s active programs and writes copy that matches the customer’s history. A merchant-side console surfaces at-risk cohorts with suggested campaigns the merchant can approve in one tap. Guardrails cap message frequency per user, respect merchant budget, and fall back to rule-based offers when model confidence is low.

What changed

On matched merchant cohorts, repeat-visit rate moved to the headline figure in the outcome block within a single quarter of rollout. Just as importantly, merchants stopped blasting every customer with the same weekly deal — message volume dropped while redemptions rose, which is the shape of retention actually working rather than discount-driven churn masking. The agent now runs as the default engagement path for new merchants onboarding to the platform.