The Digital Twin Is Only as Honest as the Master Data Underneath It

Intelligent simulation is sold as the upgrade to the planner’s scenario spreadsheet. Instead of a static model refreshed by hand each quarter, a living digital twin of the network runs continuously, and you test a disruption against it before committing real material. Gartner named intelligent simulation among its 2026 supply chain technology trends in late June. The promise is planning that has been checked against a model rather than against a guess. The promise is real. What it rests on is the part the demos skip.

The model inherits what feeds it

A simulation is only as good as the data it runs on, and the industry’s own surveys say the data is not ready. A Kaufman Rossin survey reported in July 2026 found 73 percent of mid market manufacturers still in the AI testing phase, blocked by siloed data and aging systems. At Hannover Messe in 2026, an operations executive noted that digital twins were advancing on every booth while execution at scale was working almost nowhere. Vendor material puts it more starkly still, claiming that only 1 percent of manufacturers have reached true data maturity, a figure that should be read as a sales argument for data services rather than an independent finding, but one that points the same direction as the neutral surveys. I watched the mechanism from inside. Reconciling phantom inventory and stale purchase orders and advance ship notices against what had actually arrived on the dock was months of work on the programs I planned, and until it was done, every model built on that data was confidently, precisely wrong.

The fidelity illusion

A simulation inherits the error of its inputs and hides it behind a clean interface. Call it the fidelity illusion. A twin renders inaccurate master data as a smooth, animated, executive ready picture, and the polish reads as accuracy to everyone who did not build it. The old spreadsheet had one accidental virtue: it looked uncertain. Rows of assumptions, visible caveats, a number you did not quite trust. The twin looks certain. It moves, it renders, it produces a confident plan, and the confidence is a property of the rendering engine, not of the bill of material or the lead time table underneath. This reframes what the twin is actually worth. Its value is not simulation power, which is cheap and improving. Its value is the master data discipline it forces, or fails to force. The binding constraint was never compute. It is the accuracy of the bill of material, the lead times, and the inventory records the model consumes.

The self correcting rebuttal

The serious response is that this gets the sequence backward. The twin is not something you build after the data is clean. It is how you clean the data, because running the model makes its errors visible fast, and a living twin fed by real floor data self corrects over time as the true numbers flow in. Vendors selling the large simulation suites and the newer living twin platforms make this case directly: the discipline is a feature of the tool, not a prerequisite for it.

Sometimes that holds, and it is the best case worth aiming for. But it assumes the organization funds the data remediation the twin exposes, and the common failure is the reverse. The twin is bought as the deliverable, the data work it reveals is left unfunded, and the polished model becomes the authoritative version of wrong that leadership now trusts more than the spreadsheet it replaced. The self correcting loop also depends on real floor data flowing in continuously, which lands right back on the integration problem the 2026 surveys describe: siloed data, disconnected systems, 73 percent still testing. The twin corrects itself only where the plumbing already exists, and where the plumbing already exists, the data was already better than average. The tool helps most exactly where it was needed least.

Where this holds and where it stops

The illusion bites hardest where master data governance is weak and the twin is bought as a product rather than run as a program, which describes most first deployments. It matters far less where a company already runs disciplined master data on connected systems, and there the twin genuinely adds, and in tightly instrumented new lines built with the data capture designed in. My view is the cost of the cleanup seen from inside one high volume program, which is enough to know the data work dominates the model’s value and not enough to benchmark that cost across the industry.

What follows

Before funding a twin, audit the three inputs it will run on: bill of material accuracy, lead time accuracy, and inventory record accuracy. If those numbers are unknown, the twin’s first honest output is a measurement of the data debt, and that measurement is worth buying. A twin that instead reports a confident plan on unaudited data is worth less than the spreadsheet it replaced, because the spreadsheet at least advertised its own uncertainty.

A prediction, dated so it can be checked. By 2028 the digital twin conversation moves from simulation fidelity to data readiness scoring, because buyers will have learned that the model was never the constraint and the master data always was.

A model is only alive if what feeds it is true. Feed it last year’s lead times and a bill of material nobody has reconciled, and you have not built a twin of your operation. You have built a very convincing way to be wrong on schedule.