Azirella
← Back to Autonomy

Training foundation

The Digital Twin

The digital twin is the training ground for the shared world model. It is not a separate system, it is your current planning system, replicated as a stochastic simulation. Your actual planning rules run against realistic variability, and the gap between how those rules perform and how they could perform is exactly what the AI agents learn to close.

"The digital twin concept started in manufacturing, but its greatest impact will be in supply chain. The ability to simulate your entire network in real-time changes how you make every decision."

, Michael Grieves, Research Professor, Florida Institute of Technology

How the Digital Twin Works

Your planning rules against stochastic reality, thousands of times

YOUR PLANNING RULES (DETERMINISTIC) Reorder points · Safety stocks · Lot sizes · MRP types · Procurement types Imported from your ERP, per product, per site, exactly as configured today STATIC during simulation runs against STOCHASTIC REALITY Customer demand · Supplier lead times · Inter-plant transfer times Production yield · Throughput rates · Quality pass rates · Machine availability · Changeover times Each variable randomized per entity in your network · Distributions fitted from your historical data STOCHASTIC, 50-100 Monte Carlo trials per training run produces DECISIONS PO creation · ATP allocation MO release · Inventory positioning OUTCOMES OTIF · Fill rate · Cycle time Stockouts · Excess · Cost observed by AI AGENTS LEARN WHERE YOUR CURRENT HEURISTICS FAIL

Your Planning System, Stress-Tested

Traditional planning systems use deterministic heuristics: fixed reorder points, fixed safety stocks, fixed lot sizes, fixed lead times. These rules are configured per product-site in your ERP and they work well in stable conditions.

But reality is not stable. Customer demand fluctuates. Supplier lead times vary. Machines break down. Quality issues surface. The digital twin takes your exact planning rules, imported directly from your ERP, and runs them against stochastic reality: the same demand variability, lead time uncertainty, yield losses, quality events, and machine breakdowns that your real operations experience.

The result is a precise measurement of the gap between how your current heuristics perform and how they could perform. The stockouts they didn't prevent. The excess inventory they accumulated. The late deliveries they caused. The expediting costs they incurred. This gap is exactly what the AI agents learn to close across all six decision domains.

Nine Dimensions of Variability

Every variable randomized per entity in your network, fitted from your historical data

Customer Demand

Per customer, product, site. Right-skewed, demand spikes larger than troughs. Industry-calibrated variability.

Supplier Lead Times

Per vendor, product, site. Long upper tail, disruptions cause 2.5x delays. Fitted from PO-to-receipt history.

Inter-Plant Transfers

Per lane. Tighter variability than external suppliers, internal logistics are more predictable.

Customer Delivery

Per site, customer. Moderate right skew. Fitted from actual vs. promised delivery dates.

Production Yield

Per product, site. How much usable output vs. planned input. Typical range: 90-99.5%.

Throughput Rate

Per product, site, resource. Actual vs. planned production speed. Range: 80-105% of standard.

Quality Pass Rate

Per product, site. Probability of passing inspection. Quality events cascade into supply disruptions.

Machine Availability

Per resource, site. Unplanned downtime, maintenance events. Typical availability: 80-98%.

Changeover Time

Per product pair, site. Time to switch production between products. Range: 15 min to 90 min.

Fitted from Your Data, Not Generic Assumptions

Every stochastic variable is fitted from your actual historical data when available (at least 5 observations). The system extracts the 5th percentile as the lower bound, the median as the most likely value, and the 95th percentile as the upper bound, creating a distribution that captures the realistic range of variability your supply chain actually experiences.

When history is insufficient, the system uses industry-calibrated fallbacks. Demand variability, for example, is set by product category:

15%
Staples
20-25%
Automotive / Industrial
35-40%
Seasonal / Electronics
50-80%
Promotional / Spare Parts

Each variable is instantiated per entity in your network graph, not as a single global value. A supply chain with 9 finished goods, 81 components, 27 customers, 8 vendors, and 2 plants has thousands of unique distribution instances — each reflecting the specific behavior of that entity.

Industry-Calibrated Simulation

Trials and horizon matched to your industry's characteristics

Monte Carlo Trials

Each trial is an independent replication of your supply chain operating over the full simulation horizon. More trials = more statistically reliable training data.

3PL / Wholesale50 trials
Food & Beverage / CPG50 trials
Electronics / Chemical60 trials
Automotive / Industrial75 trials
Pharma / Aerospace100 trials
Simulation Horizon

Set to 2x your industry's end-to-end supply chain lead time, enough to see one full replenishment cycle complete plus variability effects.

3PL14 days
Food & Beverage42 days
Electronics90 days
Automotive120 days
Aerospace360 days
Daily
Time bucket, weekly/monthly loses granularity
Per-site calendars
Work days from ERP, historical patterns, or country defaults
10% warmup
Initial transients excluded from training data

The Learning Sequence

Every training run follows the same eight-step sequence. This is how the digital twin transforms your ERP data into trained AI agents:

1 Load planning rules from your ERP, reorder points, safety stocks, lot sizes, per product-site
2 Instantiate distributions per entity in the network graph, thousands of unique instances from your history
3 Load guardrails, authority limits, approval thresholds, budget caps from your tenant configuration
4 Run N trials, independent Monte Carlo replications of your supply chain operating over the full horizon
5 Collect decisions + outcomes at each simulation day, what the heuristic decided and what happened
6 Compute customer service metrics, OTIF, fill rate, on-time, perfect order, cycle time for each trial
7 Score the gap between heuristic outcomes and your metric targets, where did the rules fall short?
8 Feed to AI agents, execution agents learn from individual decisions, planning agents learn from network patterns, uncertainty calibration learns from prediction-outcome pairs

The agents learn where the heuristic fails, stockouts it didn't prevent, excess inventory it accumulated, late deliveries it caused, expediting costs it incurred. The trained agents then make better decisions in the same stochastic environment.

Three Strategic Uses

The digital twin isn't just for training. It becomes a permanent strategic asset:

Policy Testing

Run 100+ simulations to test changes to inventory policies, ordering strategies, safety stock levels, or cost parameters before deploying to production. See the impact across the full range of possible futures.

Structural Testing

What happens if you add a new DC? Remove a supplier? Change your BOM? Shift production to a different plant? The digital twin models the full network effect before you make the change.

Transfer Learning

Agents trained on 100+ simulated scenarios deploy to production with learned weights, not random initialization. They arrive pre-adapted to your supply chain's specific patterns of variability.

What Gets Measured

The digital twin measures customer service outcomes, not inputs. These metrics show how well your current planning rules performed against stochastic reality, and where the gap lies:

OTIF %
On Time In Full, delivered by requested date in full quantity
Fill Rate %
Total quantity fulfilled vs. total quantity ordered
On-Time %
Orders delivered by or before requested delivery date
Perfect Order %
On time + in full + no quality issues + correct documentation
Backorder Rate
Orders with remaining unfulfilled quantity
Avg Cycle Time
Days from order placement to delivery

Metric targets (e.g., OTIF ≥ 95%) are your business goals. Metric actuals are computed from simulation results. The gap between targets and actuals is what the AI agents are trained to close.

Works With Your ERP

The digital twin imports your actual planning parameters, not generic defaults. It reads the specific reorder points, safety stocks, lot sizes, MRP types, and procurement types configured in your ERP for every product-site combination.

Supported ERP systems include SAP S/4HANA, SAP ECC, Microsoft Dynamics 365, and Odoo. The Integration layer handles the extraction and mapping automatically, including AI-enhanced interpretation of custom fields.

See Your Digital Twin

Run a simulation of your supply chain and see where the opportunities are.