Supply chain: closing the gap between knowing and reacting

The shape of the problem

A manufacturer with a multi-tier supplier base runs planning on a well-configured ERP and a capable planning team. The team is good. The process is not.

Disruption reaches the business through three channels: a supplier telling them, a delivery not arriving, or somebody reading the news. The first is rare, the second is late by definition, and the third is unreliable.

Once a disruption is known, re-planning takes days — mostly spent assembling the picture rather than deciding. Which parts are affected. Which builds those parts are in. What the alternates are. Whether the alternates are qualified. What it does to the committed dates.

Why the obvious fix does not work

The two gaps interact. A late signal followed by a slow re-plan means a long window between an event and a response, and most expedite spend is buying exactly that window back.

Improving visibility alone shortens the first gap and leaves the second. Buying a planning optimiser alone does the reverse. Neither pays for itself.

The design

Signal ingestion across the tier-two base

Shipment telemetry, supplier financial-health feeds, port and freight data, weather and regional advisories, plus a licensed news and filings feed — all resolved against the supplier master in the Cognitive Data Hub.

Entity resolution matters more here than anywhere else. A supplier appears in the ERP under one legal name, in the freight data under a facility code, in the financial feed under a parent company, and in the news under a brand. Collapsing those into one governed supplier record is what makes a signal actionable rather than merely interesting.

Predictive propagation, not just event detection

The useful output is not "there was a typhoon". It is "this typhoon affects the facility that makes your bracket, that bracket is in three builds, and the earliest impact is nineteen days out". Deterministic models compute the propagation; the agent explains it.

An agent that proposes, a planner who decides

The agent assembles the affected bill of materials, retrieves qualified alternates with lead times and pricing, models two or three re-plan options against the committed schedule, and writes up the trade-offs.

It does not commit the re-plan. A planner does, from a screen where the options are already costed. Measure the agent on how good the options are, not on how many it executes.

What to measure afterwards

• Median disruption warning lead time, against the previous baseline.
• Re-plan cycle time — and specifically the split between assembling the picture and deciding.
• Expedite spend, year over year. This is usually where the business case lives.
• Acceptance rate on proposed options: unchanged, modified, rejected. The modified bucket is the interesting one, and every modification is training signal.

Where it runs

The manufacturer's private cloud, with connectors into the ERP and the plant systems. The supplier corpus, the propagation models and the agent's tuned weights are the manufacturer's assets — which is not a legal nicety when that corpus encodes years of how a specific supply base behaves under stress.