Forecast Keeps Missing? AI for Supply Chain and Logistics Starts With the Item Master

The Fundamentals of AI for Supply Chain and Logistics

What the System Reads and What It Returns

In the deployments Synthesised.ai has scoped, three inputs carry most of the signal, and each one lives in a system the operation already licenses. Sales order history and dispatch records feed demand forecasting. Purchase orders, advance shipping notices and goods-receipt lines feed exception detection. Appointment bookings, carrier records and door-turnaround times feed dock scheduling. The item master sits underneath all three, because a model that groups two duplicate item codes as separate products forecasts each of them wrongly and the error compounds through replenishment.

Where This Sits Among Other Operational Deployments

A supply chain deployment shares its data-first sequence with every other operational build in the Synthesised.ai AI business solutions directory, and differs in which system of record carries the constraint. In a plant, the constraint sits in asset and sensor records, which is why AI for manufacturing starts with equipment hierarchies rather than item codes. In distribution, the constraint sits in the item master, the location master and the supplier master. The build sequence is the same in both cases, and the remediation work is not.

Which Master Data Objects Cap Which Use Case

Each master data object caps a specific decision, so remediation is scoped by use case rather than run as a single enterprise clean-up. The table below maps the field Synthesised.ai commonly sees fail against the use case it blocks and the check that exposes it.

Substitution and backorder flags matter more than their obscurity suggests, because unflagged constrained sales teach a model that demand equalled supply.

Why Model Choice Is Not the Constraint: Master Data Is

Item Master Fields That Bound Every Forecast Above Them

A forecast inherits the granularity of the item master, and no model recovers a distinction the data never recorded. Two item codes for one physical product halve the observed history on each, which flattens seasonality and understates peak. A missing case-pack conversion turns an accurate unit forecast into a wrong purchase order. A superseded code with live stock behind it produces a replenishment signal for an item nobody sells. In Synthesised.ai’s pre-build data profiling these three defects are among the largest sources of accuracy loss it observes, and each is fixed with reconciliation work rather than with modelling work; the specific contribution is quantified per operation during profiling.

What Distorted Inventory Costs Before Any Model Runs

IHL Group’s 2025 study estimated global retail inventory distortion at approximately USD 1.73 trillion, split into roughly USD 1.16 trillion in out-of-stocks and USD 572 billion in overstocks (an earlier IHL study projected around USD 1.7 trillion for 2024). That figure is a global retail estimate in United States dollars rather than an Australian or Singaporean market number, and it sizes the problem class rather than any single operation’s exposure. The relevance to a data-first build is direct: IHL attributes the distortion to several drivers — including empty shelves, customer service inadequacies, product location failures, supplier issues, and buying and planning errors — and a stock and demand record that does not match physical reality is a major, directly addressable contributor among them rather than the sole cause.

Governance and Privacy Obligations That Attach to Supply Chain Data

Supply chain records can include customer ship-to details and carrier driver information; where those identify an individual and are processed by a covered organisation, privacy statute applies rather than voluntary guidance alone — subject to applicable exclusions, such as Singapore’s treatment of business contact information and Australia’s small-business and other APP exemptions. The frameworks below fall into two groups: applicable legislation — Singapore’s PDPA and Australia’s Privacy Act 1988 — which binds where it applies, and voluntary governance standards and tools — ISO/IEC 42001, the NIST AI Risk Management Framework, DISR Guidance for AI Adoption and IMDA AI Verify — which apply where adopted contractually or through another binding requirement. Synthesised.ai names the set below in scoping for supply chain builds across Singapore and Australia.

Is a Better Model Able to Compensate for Poor Master Data?

Not on its own. A stronger model fits the recorded history more closely, and a record that misdescribes the physical goods is fitted more closely to the wrong thing, so where data quality is the binding constraint a model upgrade is bounded by the noise the data contains. De-duplicating an item master removes that noise at source; model specification and configuration can also matter, so the two are weighed by benchmarking on the specific deployment. Synthesised.ai sequences the two accordingly, with data profiling and remediation scoped before any model selection conversation begins.

Forecasting, Exception and Dock-Scheduling Use Cases

Demand Forecasting on Transactional History

Demand forecasting learns from order and dispatch history, and the first build step reconstructs true demand from sales that supply constrained. Sales records report what the operation was able to ship rather than what customers wanted, so a period of stock-out reads as a period of low demand unless backorder and substitution flags correct it. An initial historical reconstruction feeds the baseline, and the same treatment is built into the ongoing pipeline so that new stockouts, substitutions and censored observations are reconstructed as they occur, keeping the training history a description of demand rather than availability.

Exception Handling Across Orders, Shipments and Invoices

Exception handling ranks the orders that need human intervention today, which turns a coordinator’s morning from a search into a queue. The model scores each open order against expected receipt dates, quantity variances and invoice mismatches, and hands the coordinator a list ordered by consequence. The same routing logic appears in the wider account of AI workflow automation, where the automated layer clears the routine cases and escalates the rest with the evidence attached.

Dock Scheduling and Yard Slot Allocation

Dock scheduling allocates doors and time slots against inbound loads, and which system and data objects control it is architecture-dependent — dock-door capability, appointment and yard state, arrival uncertainty, labour and equipment, load attributes and carrier constraints can sit across the WMS, TMS and YMS rather than in a single master. A model that assigns a chilled inbound to an ambient door produces a schedule the yard ignores. Dock-door capability (distinct from storage-location attributes), honest turnaround histories and current appointment and yard data are among the inputs that make the allocation actionable, and where they live depends on the operation’s system architecture.

Which Use Case Goes First in a Data-Constrained Operation?

Exception handling goes first when master data quality is poor, because it tolerates imperfect records better than forecasting does. An exception model compares transactions against each other and surfaces mismatches, which means a defective record often becomes a useful exception rather than a silent error. Forecasting inherits every defect as accuracy loss. Sequencing exception handling first delivers a measurable result while the item master remediation runs in parallel.

Verifying Forecast and Service-Level Gains

Baselines Fixed Before the Model Ships

A baseline agreed after go-live is a negotiation rather than a measurement, so every figure above is fixed in writing during scoping. Forecast error uses the operation’s own existing definition where one exists, because substituting a different error metric mid-programme makes the comparison meaningless. Where no definition exists, one is written into the scope document and signed by the planning owner before the first model runs.

The Frozen Evaluation Set and the Release Gate

Every release passes a frozen evaluation set before it reaches planners, and the set is versioned alongside the model. The evaluation set holds out a period of real history the model has never trained on, including at least one demand peak and one supply disruption. A release that degrades against the held-out period does not ship. This gate produces the same monitoring records the ISO/IEC 42001 clauses and the NIST AI Risk Management Framework measurement functions ask an organisation to keep.

Troubleshooting a Supply Chain AI Deployment

Forecast Accuracy Falls Away After Go-Live

Accuracy decay after go-live often traces to master data changing underneath a model that trained on the old records, but drift, a changed demand regime, model misspecification or an integration failure are also plausible, so the cause is confirmed by log-based root-cause analysis before attribution. New item codes arrive without the attributes the model reads, a supplier lead time is edited without history, or a location is repurposed without updating its storage type. The fix is a change-control routine on the master data objects the model depends on, with monitoring that alerts on attribute completeness rather than on accuracy alone. Accuracy monitoring reports the symptom several weeks after the cause.

Planners Override the Recommendation and Keep Their Spreadsheet

A high override rate is often a data-credibility signal rather than a training problem, though incentives, workflow design and genuine model error can also drive it; the reason codes attached to overrides help name the actual defect. Planners override recommendations that reference items they know are discontinued, or quantities that ignore a case-pack rule they apply by hand. Capturing a mandatory reason code on each override turns planner scepticism into a remediation backlog. The override rate falls as that backlog clears, and it falls without any change to the model.

The Pilot Works and the Rollout Stalls

A pilot that succeeds on one site and stalls across the network is often meeting master data that was never harmonised between sites, though integration failures, capacity constraints and site-level workflow differences can also stall a rollout, so a rollout post-mortem confirms which applies. Item codes diverge, unit-of-measure conventions differ, and location naming follows local habit. The pilot proved the mechanism and the rollout is exposing the data estate. Scoping the harmonisation as its own workstream, with a named owner per site, is what converts the pilot result into a network result.

Why Do Supply Chain AI Projects Stall Without a Written Strategy?

A written strategy fixes which decision the model improves, who owns that decision, and what evidence closes the question, and a project missing all three has no definition of completion. In Synthesised.ai’s experience, deployments frequently run ahead of that definition. Synthesised.ai writes the decision, the owner and the evidence into the scope document before the build starts, which makes the completion test explicit from the first sprint.

What Do the Data-Quality Findings Mean for an Australian Operation?

The published figures on this page come from global and United States research rather than from Australian samples, and they size the problem class rather than any local benchmark. The PwC survey of 767 executives is a United States sample published on 23 April 2026. The IHL Group inventory distortion estimate is a global figure stated in United States dollars. An Australian or Singaporean operation gets its own numbers from its own systems during profiling, which is the only baseline a verification programme accepts.

Scope a Data-First Build