The hypothesis
Regional distributors face a maddening paradox: empty shelves for fast-moving items while slow-movers pile up in the warehouse. A 280-employee industrial distributor was running 18% stock-out rate on A-class SKUs while carrying 147 days of inventory on C-class items. Their demand planning ran monthly cycles, but customer patterns shifted weekly.
We believed an agent could eliminate this paradox by continuously sensing demand signals across all channels and automatically triggering replenishment adjustments before stock-outs occurred. The manual cost: two planners spending 28 hours weekly on emergency reorders and overstock analysis, plus $340K annually in lost sales from stock-outs.
What we tested
We built a 4-component demand sensing agent that processes real-time inventory movements, customer behaviour patterns and supplier lead times. Unlike monthly planning cycles, this system evaluates every transaction as it happens.
Signal Aggregator: Ingests order patterns, inventory movements, customer enquiries and supplier confirmations from the WMS and customer portal. Outputs structured demand signals with velocity indicators and pattern classifications.
Velocity Tracker: Analyses 90-day rolling patterns for each SKU, detecting acceleration or deceleration trends. Identifies when an item's velocity diverges from its historical classification. Outputs velocity change alerts with confidence scores.
Threshold Calculator: Combines current velocity with supplier lead times and safety stock requirements to determine optimal reorder points. Adjusts thresholds based on seasonal patterns and promotional calendars. Outputs dynamic min/max levels for each SKU.
Action Router: Compares current stock levels against calculated thresholds and routes recommendations. Auto-generates purchase orders for routine replenishment, escalates unusual patterns to planners, flags potential overstock for markdown consideration.
Test dataset: 4,200 active SKUs across 8 product categories, processing 1,800 daily transactions over 12 weeks. The system evaluated every inventory movement within 15 minutes of WMS posting.
The architecture
WMS INTEGRATION ZONE DEMAND SENSING AGENT PIPELINE
┌─────────────────────┐
│ Warehouse Mgmt │
│ • Inventory moves │──────┐
│ • Receipts/picks │ │
│ • Cycle counts │ │ [1] Signal [2] Velocity [3] Threshold
└─────────────────────┘ ├───► Aggregator ────► Tracker ────────► Calculator
┌─────────────────────┐ │ │ patterns │ │ trends │ │ min/max │
│ Customer Portal │ │ │ velocity │ │ alerts │ │ levels │
│ • Order enquiries │──────┘ └──────────┘ └─────────┘ └─────────┘
│ • Repeat patterns │ │
└─────────────────────┘ │
▼
┌─────────────────────┐ [4] Action Router
│ ERP Finance/AP │◄──────────────────────────────────────────── │ auto-reorder │
│ • Auto PO creation │ ┌─────────────────────────────── │ escalations │
│ • Approval routing │ │ └──────────────┘
└─────────────────────┘ │ │
│ │
┌─────────────────────┐ │ MATCH: Normal velocity, │
│ Planning Team │◄─────────────┘ standard reorder patterns │
│ • Exception review │ │
│ • Pattern analysis │◄───────────────────────────────────────────────────────┘
│ • Override approval │ EXCEPTION: Velocity anomalies,
└─────────────────────┘ unusual demand patterns
What worked
The agent eliminated the stock-out/overstock paradox by sensing demand changes 3-4 weeks before traditional planning cycles would detect them.
- Velocity detection caught demand shifts 23 days earlier than monthly planning cycles
- Auto-reorder accuracy: 96% of generated purchase orders required no human modification
- Overstock identification flagged 340 slow-moving SKUs for markdown before they became dead stock
- Exception escalation: only 4% of decisions required planner intervention
- Customer service impact: 89% reduction in "sorry, we're out" responses
- Cash flow improvement: $127K reduction in excess inventory carrying costs over 12 weeks
- Seasonal item velocity tracking required 6 weeks of learning before reliable predictions
- New product introductions still needed manual threshold setting for first 30 days
- Supplier lead time changes took 2-3 weeks to propagate through threshold calculations
- Promotional demand spikes occasionally triggered false shortage alerts
What we learned
- Real-time sensing beats periodic planning. The agent's continuous monitoring caught velocity changes that monthly planning cycles missed entirely. Demand patterns shift faster than planning schedules.
- Dynamic thresholds prevent both problems. Static reorder points cause the stock-out/overstock paradox. The agent's calculated thresholds adapted to each SKU's actual velocity, eliminating the guesswork.
- Exception routing preserves planner expertise. The agent handled routine replenishment automatically but escalated unusual patterns to human planners. This preserved institutional knowledge while eliminating administrative work.
- WMS integration amplifies the agent's impact. The real value emerged when the agent could automatically post purchase orders to the ERP and update inventory parameters in the warehouse management system. The agent is the brain, but the enterprise systems are the nervous system that executes the decisions.
Potential for our clients
Experiment status
The demand sensing agent is production-ready for distributors with established WMS systems and at least 6 months of transaction history. The velocity tracking component requires 4-6 weeks of learning before reliable threshold calculations. Integration complexity depends on ERP purchase order automation capabilities.
Operations directors at regional distributors should evaluate this approach if they're spending more than 20 hours weekly on manual replenishment planning or carrying more than 120 days of inventory while experiencing regular stock-outs. Typical engagement scope is a Workflow Sprint to map current planning processes and design the agent integration with existing warehouse management systems.