A Worked Example from groa°: Autonomous VIP Decay Intervention
By Polly Barnfield, OBE, CEO of Maybe*
Theory is cheap. Here is what happens, end to end, when a VIP customer starts to slip, and how the groa° agents, built on Maybe*, intervene before churn hardens.
Every edition of Scaling Up With Agents includes a worked example, because architecture diagrams are not the same thing as a system doing the job. The example in this post comes straight from Chapter 7 of Retention-First Growth®, expanded into the engineering reality of how it runs on the platform we are building with groa°.
The scenario: a Loyalty-orbit customer, top 10% by lifetime value, missing their expected 45-day repurchase window. In a manual environment, this customer is invisible until their dormancy duration crosses a static threshold weeks later. By which point the recovery window has closed and recovery rates have collapsed from the 20 to 40% the research describes for early intervention into single digits.
Here is what the same scenario looks like inside the groa° agentic platform.
Hour 0: signal detection
The customer's last purchase was day 0. The Loyalty-orbit benchmark for their cohort and their personal repurchase cycle says the next purchase should fall within 35 to 45 days. The continuous observation layer is tracking this against their Customer Energy Profile™. Recency, frequency, monetary value, engagement signals across email and SMS, loyalty tier velocity.
On day 38, the customer's email engagement starts to thin. Open rate on the last three campaigns has dropped from their personal baseline. They have not visited the site in 12 days, against a baseline of every 6 to 8 days. Each of these signals on its own would not warrant action. False positives are expensive in retention. Combined, they push the agent's confidence-weighted risk signal above threshold.
This is the moment manual systems miss. None of these individual data points would surface in a weekly review. The pattern only resolves when you are watching all of them, against this specific customer's baseline, continuously.
Hour 0 to 24: state interpretation
The agent does not yet know whether to intervene. It knows there is a high-confidence risk signal. The question is which response, and when.
This is where the orbit governance layer does its work. The customer is in the Loyalty orbit, so the relevant decision policy is Loyalty-specific, not a generic winback flow. The Customer Energy Profile™ shows their fatigue signals are concentrated on promotional content. Their loyalty tier history shows they responded strongly to early-access mechanics in the past. Their last three purchases were full-margin, not discount-led.
The agent reads this and reaches a conclusion that a workflow could not. This customer is not price-sensitive. Pushing a discount would erode margin without addressing the actual decay driver, which is engagement fatigue with broadcast content. The right intervention is recognition-led, not discount-led. Specifically: an early-access invitation to an upcoming limited-edition launch, framed as a tier benefit, delivered through email rather than SMS. Because SMS at this energy level would feel like pressure, not privilege.
Hour 24 to 48: governed intervention
The intervention executes through the existing Klaviyo infrastructure. Nothing about this is exotic. Klaviyo is, as the book makes clear, an excellent execution engine. What changes is the upstream decision: which customer, which content, which channel, which moment, governed by orbit-specific logic rather than a calendar.
Sent at the time of day this customer historically opens email, with content adapted to their inferred preferences from their loyalty interaction history, with a single clear action aligned to their lifecycle position.
Hours 48 onwards: closed-loop learning
The customer opens. Clicks. Adds to cart. Purchases on day 41, within the original repurchase window, with no margin erosion, with their loyalty tier velocity reinforced rather than reset.
That outcome feeds back into the policy layer. For this cohort, at this energy profile, an early-access mechanic at the 38-day mark with a 24-hour follow-up window produced a positive outcome. The next customer who matches this pattern will be handled with slightly tighter confidence thresholds. The pattern that did not work, the discount-led intervention the agent rejected, is reinforced as the wrong move for this profile.
Multiply this across thousands of Loyalty-orbit customers, hundreds of brands, every day, and you have the operating system Emma described in the book. Continuous observation. Orbit-aware reasoning. Goal-directed intervention. Bounded learning. None of it possible to sustain manually. All of it possible inside an agentic platform built for it.
What this is not
It is not a black box. Every decision is auditable. The brand and the groa° team can see why each intervention fired, what signals triggered it, and what alternatives the agent considered. Governance-first means the human team retains authority over the rules. They tune the orbit policies, set the economic guardrails, and can override anything they want.
It is not a replacement for the strategy work. The agent does not decide what the brand stands for, what the loyalty programme rewards, or what tone the messaging should take. Those are human decisions, encoded into the system upfront, and refined as the brand evolves.
It is the layer that makes top-decile retention performance durable at scale. Which is what the book has been arguing for from page one.
This is part 3 of 4 in the groa° edition of Scaling Up With Agents. Final post in this edition: how this changes the role of retention teams, and what cognitive partnership actually looks like when it works.
The full retention methodology is in Emma's book. Read it here.
Want to see how this would play out in your own brand's data? Book a call with us.
