Trust & Control

AI Model Governance

Every model that informs a decision is registered, versioned, monitored for drift and approved by a named business owner. Humans approve every action; outcomes are fed back to retrain models.

Model lifecycle

1. Development1

Built & back-tested by Data Science

2. UAT1

Shadow mode, business validation

3. Approved1

Signed off, awaiting go-live

4. Production6

Serving live decisions

5. Monitoring1

Live with an active drift alert

Learning loop · closing the feedback cycle

Demand Forecasting: drift detected → challenger model → business validation → outcome feedback

How the West Java stockout became a better model

Learning

1Drift alert

v3.4.1 data drift PSI 0.19 (Moderate). Promotion-depth distribution shifted with the Care Fest 15% discount.

2Root cause

Planning forecast assumed +10% uplift; actual promo demand was +41%. Replenishment was sized on the wrong number.

3Challenger in UAT

v3.5.0 adds promo depth, display & distributor stock features. Promo-period WMAPE 31% → 12%.

4Business validation

ACT-002 · promote v3.5.0 to ProductionIn review

5Outcome feedback

3 completed actions compared predicted vs actual impact; calibrations fed back into training data.

Forecast error (WMAPE)

Back-test on history · lower is better

v3.4.1v3.5.0

v3.4.1 data drift · 12 weeks

PSI · dashed = moderate (0.10) and significant (0.20)

Action outcomes · predicted vs actual

Completed actions feed model calibration

PredictedActual

ACT-090Transfer 18K units Jakarta DC → Medan DC. · Revenue protectedPredicted IDR 420MActual IDR 390M−7%Model calibration updated: transfer lead time +0.5 day for Sumatra lanes.

ACT-091Replace valve set during planned stop. · Downtime cost avoidedPredicted IDR 380MActual IDR 410M+8%Failure signature confirmed on removed valve; label added to training set.

ACT-092Win-back journey with sunscreen sample. · Revenue retainedPredicted IDR 310MActual IDR 270M−13%Response lower in GT-first consumers; churn model feature 'first_channel' re-weighted.

Model registry

Responsible AI controls

Enforced

Human-in-the-loop

No AI recommendation executes automatically. Every action requires a named approver in the Action Center; P1 actions need director sign-off.

Enforced

Audit trail

Every recommendation, simulation, approval and outcome is time-stamped with actor and model version, retained 7 years.

Enforced

Explainability

Each prediction ships with drivers, evidence and data lineage; business users see why, not just what.

Enforced

Data privacy

All data in this prototype is synthetic. Consumer IDs are hashed; no names, phone numbers or addresses are used in modelling.

Monthly

Bias monitoring

Consumer models checked monthly for region & channel parity (max gap 3.1 pts for churn). Offer depth capped equally across segments.