The word is used for three different things, and only one of them is a digital twin.
A digital model is a virtual representation with no automatic data link — a CAD assembly, a plant layout drawing. You update it by hand. A digital shadow takes a one-way data feed from the physical asset, so it reflects reality but cannot act on it. A digital twin closes the loop: data flows from the asset to the model, and decisions flow back. Change something in the twin, and you can push the change to the machine.
Most products sold as digital twins are digital shadows. That is not a scandal — a shadow is useful, and plenty of the outcomes below came from one. But it changes what you should expect, and what you should pay.
What digital twins are actually used for in manufacturing
Our corpus holds 125 published manufacturing case studies that use digital twin technology. Here is what they were built to do:
| What the twin was used for | Deployments | Share |
|---|---|---|
| Process optimization | 63 | 50% |
| Quality control & inspection | 27 | 22% |
| Supply chain optimization | 12 | 10% |
| Safety & compliance monitoring | 11 | 9% |
| Everything else | 12 | 10% |
| Predictive maintenance | 2 | 2% |
That last row is the finding worth carrying into a vendor meeting. The popular framing of a digital twin — a living replica of your machine that warns you before it breaks — accounts for two of 125 documented deployments. Predictive maintenance in this corpus overwhelmingly runs on sensor models and vibration analytics, not on twins. If a vendor's pitch is "digital twin for predictive maintenance", they are selling into the thinnest evidence base in the category.
By industry, the deployments concentrate in Industrial Machinery (29), Automotive (24), Food & Beverage (15), Metals & Mining (13) and Pharmaceuticals (10).
The outcome digital twins actually deliver is time
Of the 125 deployments, 46 report a quantified outcome. Read them together and one unit dominates: time before production, not uptime during it.
- Kasa cut baggage-handling system commissioning from 4–6 weeks to 6 days using an emulation twin of the system (case study).
- ECM Technologies cut commissioning time 50% on automotive heat-treatment equipment (case study).
- GROB reported a 30% reduction in CNC programming time, and Ural Locomotives a 40% reduction in time-to-manufacture, both from twins built during design rather than after installation (case study).
The pattern holds across the rest: design time, programming time, commissioning time, lead time. A twin earns its cost by letting you make mistakes in software, at the point where a mistake is a rebuild of a model instead of a rebuild of a cell.
Worked example: what "6 weeks to 6 days" is actually made of
Commissioning a materials-handling system on site means running the real conveyors against the real PLC code, finding the logic faults, fixing them, and running again — with the line down and integrators on the clock. Kasa built an emulated twin of the mechanical system and ran the same PLC code against it in the office. The faults surfaced against the model. What remained on site was verification.
The saving is not that the twin optimised anything. It is that the debugging moved off the critical path. That is the honest mechanism behind most digital twin ROI, and it tells you when a twin is worth building: when downtime during commissioning or changeover is expensive, and when the logic you are testing is complex enough to hide faults. A single-station cell with twelve lines of ladder logic does not need one.
What it takes to run one
A twin is only as current as its data feed, which is where most of the cost sits. You need tag-level data off the control system at a usable rate, a model whose fidelity matches the question you are asking, and someone who owns the model when the plant changes. Plants change constantly; a twin nobody updates becomes a digital model again within a year, quietly.
Fidelity is a decision, not a maximum. A twin built to test control logic needs accurate timing and sequencing and can ignore thermodynamics. A twin built to predict product quality needs the physics and can ignore the cosmetics. Vendors will happily sell you fidelity you have no question for.
How AI changes the picture
Two ways, and they are not the same.
Surrogate models. A physics-accurate simulation can be too slow to run inside a control loop. Training a machine-learning model on that simulation's outputs produces a fast approximation that runs in real time — the twin's answers at a fraction of the compute. This is where "AI digital twin" usually means something specific.
Model calibration. Real equipment drifts away from its design spec: bearings wear, heat exchangers foul, tolerances stack. ML fitted to live sensor data keeps the twin's parameters tracking the machine as it ages, which is the difference between a twin that stays useful for five years and one that quietly stops matching the plant.
Both are additions to a working data pipeline. Neither substitutes for one — which is why the twins that deliver in the corpus are overwhelmingly at companies that already had their machine data in order.
Where to go next
- The digital twin technology page lists every documented deployment in the corpus, filterable by industry.
- If you are evaluating platforms, the vendors that show up most in these deployments are ranked with their evidence on the manufacturing operations management software hub.
- For the broader process-optimization outcomes these deployments sit inside — medians, ranges and sample sizes — see the process optimization benchmarks.
- Eaton's Changzhou plant is the clearest example of a twin used for ongoing operations rather than commissioning: 39% lead-time reduction and 50% operational efficiency increase (case study).
AI Roadmap — Eight workflows in your plant where AI would pay back this quarter. We know which. 12 pages, in 72 hours.