How Retention Intelligence works
Five stages, one continuous loop. Closed-loop with attribution.
Detect → Prioritize → Activate → Act → Attribute. Every rep action feeds back into the model — sharper predictions next week, sharper still next month. The deep-dives below explain the science behind two of them.
01
Detect
Survival-analysis AI/ML predicts revenue at risk at the SKU level — and rolls it up to account-level Annual Revenue at Risk. The model runs three times a day across every account and every SKU, learning each customer's unique ordering pattern per product and separating real churn from normal variation.
SKU + account level · 3× daily refresh
02
Prioritize
Agents rank Next Best Action — ten accounts a day, by Annual Revenue at Risk. Not a static report. Re-ranked every morning on new signals.
10 accounts · ranked by ARR at risk
03
Activate
Agents fire across every customer surface — daily rep digest, order entry, customer-service inbox, CRM tasks, e-commerce reorder prompts. Wherever the customer is, the agent is.
Omnichannel · digest · order entry · CRM · CS · e-com
04
Act
Rep dismisses what's not real, snoozes what's not urgent, pushes what matters to CRM and owns it. Every action captured and attributed back.
One-click · dismiss · snooze · push
04
Attribute
System measures recovery. Specific products, specific customers, specific dollars recovered — not estimates. The closed loop. Proof you can show the board.
Closed-loop · per-SKU · per-customer
Stage 01 · Detect
Drift vs. Real Churn
A customer who orders every 90 days and is on day 75 looks identical to a churned customer to a threshold rule. That's why distributors haven't built this internally — threshold-based reorder reports flag every late account, and the queue is too noisy to work. SETVI uses survival analysis (the statistical technique medical researchers use to predict time-to-event outcomes) running three times a day on every account and SKU. The model learns each customer's individual cadence per product, so only patterns that have actually broken for that specific customer surface to the queue. Thresholds on aggregates ask "is this account late?" Survival analysis asks "is this account late for them?" The hardest problem in retention, solved.
Stage 04 · ACT
Rep-Trained Refinement
Dismissals teach which predictions to trust. Snooze patterns reveal optimal contact windows. Six months in, predictions sharpen materially — based on how your reps actually work the list. The system gets harder to compete with over time, not easier.
Self-improving
Drift vs. churn — concrete example
Customer X orders gloves every 8.5 days, is 3 cycles late, $17K/year at risk, #3 on today's action list. A threshold rule would flag this customer at day 14 along with everyone else. The model knows their cadence specifically — and surfaces them at day 25, not day 14, because that's when the pattern actually breaks for this customer.
.png)
