Sentinel Intelligence · AI-Native Architecture

Player Churn Modeller

Score every player every day, intervene before the churn date — collect → score → route by tier → intervene → verify the save. Mass and Premium run on VELA-drafted automated journeys; VIP is always a human conversation, briefed by the engine, never a bot.

Tier-routed: auto · assisted · human Powered by VELA RG guardrails enforced upstream Synthetic demonstration
Scenario
01

Command view

One screen, every at-risk dollar · synthetic demo
Value at risk — next 30 days
Value at risk = Σ flagged players: P(churn) × 6-month expected NGR. Saved MTD = verified saves × realised 6-month value, net of offer cost.
Decision loop — today's scan
RG suppression runs before routing: self-excluded, cool-off, RG-watch and opt-out players are removed from the contact pool — no message, no offer, no exception.
Advisory — what the engine recommends today
Each recommendation carries its evidence and expected impact — auditable, never a black box.
02

The churn model

Gradient-boosted survival model · retrained weekly · synthetic demo

Risk score distribution — active base

Intervention threshold at score ≥ 65. Everything right of the line enters the queue.

What drives the score

Top signals, by model weight.

03

Tier routing — who gets a bot, who gets a human

The core design decision · synthetic demo
Fully automated

Mass

flagged today · auto-queued
  • VELA drafts per-player email / SMS / push from the churn driver
  • Sent automatically in each player's optimal window — no agent touch
  • Offer capped by tier economics; A/B control group held out
  • Frequency cap 1 contact / 72h across all channels
Agent-assisted

Premium

flagged today · awaiting agent action
  • VELA drafts the response; CRM agent approves in one click or copy-pastes into any channel
  • Agent may edit before send — every edit trains the drafts
  • SLA: actioned within 4 working hours of flag
  • Escalates to host if two touches go unanswered
Human only

VIP

flagged today · host assigned
  • Never automated. A VIP at risk is a phone call, not a template
  • VELA prepares the host briefing: context, talking points, approved gestures
  • Host logs the outcome; engine verifies the save at day 30
  • SLA: host contact within 24h of flag
04

Intervention queue

Click a player to open the VELA response · synthetic demo
Est. net save = P(churn) × 6-mo NGR × tier save-uplift − offer cost. Queue ranked by net save value, not by score — the engine spends effort where the dollars are.
05

VELA response studio

Copy-paste for agents · auto-send for mass · briefing for hosts
06

Save performance

Verified at day 30 vs held-out control · synthetic demo

Weekly saved value — trailing 12 weeks

Net of offer cost, verified saves only.

Net saved valueOffer cost

Save rate by tier — vs control

Uplift = treated save-rate − control save-rate. Control groups are held out per tier every cycle; a retention programme that can't beat its control is a discount programme.
07

Responsible-gambling guardrails

Enforced in code, upstream of every message
Suppressed from today's contact pool
Suppression is absolute: a suppressed player cannot be selected, drafted for, or contacted by any channel — including manually from this screen.

The rules the engine cannot break

Self-excluded and cool-off players are invisible to the retention engine — removed at ingestion, not at send.
📉
Loss-chasing signals suppress offers. A churn driver of "big loss + rapid redeposits" routes to an RG check, never to a win-back bonus.
🕐
Frequency caps: max one retention contact per 72 hours across all channels; quiet hours respected per jurisdiction.
💬
No urgency or pressure language. VELA's drafts are screened: no "last chance", no countdowns, no loss-framing. Plain-language bonus terms in every message.
⚖️
Jurisdiction-aware: bonus rules, opt-in requirements and marketing-consent flags applied per licence; every send logged for the regulator.
08

Live support · VELA copilot

VELA drafts the agent reply, screens it, agent sends · synthetic demo
Inbound player message
VELA draft for agent
Every VELA draft is screened for urgency, countdown and loss-framing language before the agent sees it. RG-flagged messages are replaced with a safer-gaming response and held for human review — they never auto-send.