PEKO case study — what an AI retention rollout looks like in practice

This is a composite rather than a single venue: the pattern below repeats closely enough across independent F&B rollouts that describing one venue would be less honest than describing the shape. Figures are outcome bands observed across deployments, not a guarantee for any particular site.

Starting position. A typical venue arrives with a points programme that technically exists: a POS module or a stamp card, single-digit identification of transactions, no automated messaging, and no way to say who is at risk. Revenue is flat, the owner believes retention is fine because the regulars they can see are still coming, and nobody is counting the regulars who are not.

Phase one, weeks 1–2: enrolment only. Counter and table QR, receipt capture as a fallback, a first-scan reward claimable on the next visit. Deliberately no messaging in this phase — the point is to build an identified base and a clean baseline. Identification typically moves from under 10% of transactions to 40–60% within a fortnight, and this single change is responsible for most of the eventual result, because everything downstream is multiplied by it.

Phase two, weeks 3–4: welcome and birthday. Two automations, both low-risk and both well received. The welcome flow converts first visits into second visits inside a week; the birthday flow is consistently the highest-ROI single message in the stack. No win-back yet, because the cadence model needs data.

Phase three, weeks 5–8: cadence-based win-back with a permanent 10% holdout. The agent computes each customer's own visit rhythm, flags deviation at 1.5× and 2×, drafts the message and queues it for approval. Compared against the venue's previous 30-day-rule blasts, cadence triggering typically converts two to three times better while sending materially fewer messages.

Phase four, weeks 9–12: the first matched-cohort review. Members versus matched non-members at 90 days, plus messaged versus held-out among lapsers. The bands that come out of this review across deployments: 10–18 points of 90-day repeat-rate improvement, 8–15 hours a month of marketing work removed from the owner's plate, and a win-back holdout gap that makes the programme's incremental contribution explicit rather than assumed.

What consistently goes wrong when a rollout underperforms: enrolment stalls below 25% because the QR is placed where nobody waits, or the venue switches on all automations at once and cannot attribute any of the movement. Both are sequencing failures rather than product failures, which is why the phase order above matters more than any individual setting.