Key takeaway
- Digital twin in manufacturing explained for enterprise leaders: ISO 23247 and NIST standards, use cases, limits and a sourced 3 to 5 year outlook.
A digital twin in manufacturing is a virtual model of a machine, line or factory that is kept in step with the real one by live data, so engineers can test changes, predict faults and plan production before touching the physical plant. The Digital Twin Consortium describes a digital twin as "an integrated data-driven virtual representation of real-world entities and processes", and ISO 23247 gives manufacturing its own framework for building one.
For enterprise leaders the useful question is not whether twins work but where they pay back and what they need underneath them: trusted data, agreed standards and clear ownership of decisions. This guide covers how a factory twin works, where it is used, its limits, and what is likely to change over the next three to five years. For the wider concept see our digital twin technology guide, and for our delivery offer see digital twin consulting.
What is a digital twin in manufacturing?
McKinsey defines a factory digital twin as a "real-time virtual representation of the factory" that supports faster and more cost-effective decisions (McKinsey, January 2024). ISO 23247-1:2021 sets out the overview and general principles of a digital twin framework for manufacturing, and ISO 23247-2:2021 adds a reference architecture with domain, entity and functional views. Both were published in October 2021.
Smart manufacturing is the wider setting. Machines connected through the Industrial Internet of Things send operating data that analysts and engineers use to spot faults early and to plan work. A twin adds a model on top of that data, so the question moves from "what happened?" to "what will happen if we change this?".
How does a digital twin work in a factory?
A twin starts with applied data science: the physics or mathematics of a machine or process, plus its operating data, are turned into a model that simulates the original. The steps below follow the usual build order.
| Step | What happens | What the decision maker gets |
|---|---|---|
| 1. Collect and model | Sensors on the real asset feed a virtual computer model of it. | A model tied to real operating conditions |
| 2. Simulate in real time | The twin mirrors the asset and shows performance, efficiency and likely problems. | Early warning before a fault stops a line |
| 3. Scale the detail | A twin can be simple or complex depending on the data available; more data gives a more accurate, responsive simulation. | A way to start small and add detail where it pays |
| 4. Prototype and test | A twin can sit beside a physical prototype to validate a design, or stand alone to predict how the product will behave in production. | Design flaws found before manufacturing begins |
Where are digital twins used in manufacturing and beyond?
Twins are used well beyond the factory floor, which is why the same ideas keep returning in different sectors. The table lists the applications the original guide covered and where each fits.
| Area | What a twin is used for | When it fits |
|---|---|---|
| Production lines and plants | Testing schedules, sequencing and process changes on the model first | Complex lines where trial and error on the floor is costly |
| Power generation and automotive | Testing problems such as exhaust behaviour, corrosion resistance of offshore wind turbines and racing-car efficiency | Assets that are expensive or hard to test physically |
| Smart cities | Virtual models that guide planning decisions | Planning for sustainability across many systems |
| Healthcare | Surgery training, hospital workflow and staffing models, and tracking where infection may spread | Process improvement; not a substitute for clinical judgement |
| Retail and disaster management | Modelling customer experience; infrastructure, climate monitoring and emergency-response plans | Scenarios that cannot be rehearsed in the real world |
In McKinsey's June 2022 survey of 75 senior industrials leaders, 86 percent said a digital twin applies to their organisation, 44 percent had implemented one and 15 percent were planning to. This is one firm's small survey, so treat it as a signal and not a market size.
What are the benefits of a digital twin in manufacturing?
- Remote monitoring: an extensive physical system is hard to see as a whole; a twin can be reached from anywhere, so teams can monitor and control performance remotely.
- Faster risk assessment and production planning: products and planned processes are validated before they exist, so errors are found before manufacturing starts.
- Reliability under disruption: engineers can disturb the model, observe the reaction and work out mitigations without risking the line.
- Collaboration: shared, always-available system data lets teams work from one picture.
- Predictive maintenance: sensor data analysed in the twin points to problems early, so maintenance can be planned more accurately and downtime avoided.
McKinsey reports case examples: a schedule redesign at an assembly plant produced a "5 to 7 percent monthly cost saving", and optimised sequencing on a production line cut total processing time by "about four percent". These come from unnamed client cases and McKinsey does not explain how they were measured, so use them as illustrations only.
What are the challenges of implementing a digital twin?
NIST calls the lack of standards "a major barrier": there are no common rules yet for vocabulary, design, interoperability, trustworthiness or validation methods. It adds that building a twin correctly is hard, most of all for small and medium-sized enterprises, and that models need validation with quantified uncertainty, which relies on measurement science still being developed.
McKinsey says development can take two to three months or more than a year, depending on team experience and factory complexity. The practical risks for a leader are a twin that drifts from the real plant, data that nobody owns, and decisions taken on a model that has not been validated.
| Trade-off | What to do | When a twin is not the right first step |
|---|---|---|
| Data quality | Fix sensor coverage and data ownership first | Machines are not instrumented and no one owns the data |
| Model validity | Validate against real runs and record the uncertainty | No one can say how accurate the model needs to be |
| Interoperability | Prefer open, documented interfaces and standards such as ISO 23247 | The plan depends on one closed toolchain you cannot replace |
How do digital twins connect to edge AI and agentic workflows?
A twin is most useful when it is connected to the systems that act. Edge AI runs models close to the machines, which our guide to edge AI in smart factories covers, and agentic AI in manufacturing looks at agents that plan and act inside governed limits. For Microsoft's own twin platform, our review of the Azure Digital Twins platform covers its architecture and status.
Microsoft describes Azure Digital Twins as "a platform as a service (PaaS) offering that enables the creation of twin graphs based on digital models of entire environments", with models written in the Digital Twins Definition Language (DTDL) (Microsoft Learn, page last updated 28 January 2025). Where an agent acts on twin output, the decision needs a record of the context and policy behind it. ElixirData's Context OS, according to its site, keeps governed context, versioned policy and decision traces in one decision layer above AI agents, and AgentAnalyst lets teams ask questions of enterprise data in plain language. For physical AI that runs at the edge, NexaStack describes itself as an agentic operating platform for physical AI.
What will change in digital twins in manufacturing over the next 3 to 5 years?
The sources below describe work under way. The "likely direction" column is our reading of them, not a forecast from the source, and none of it is a promise of dates.
| Area | What the source says today | Likely direction |
|---|---|---|
| Standards | ISO 23247-1 and -2 were published in October 2021. NIST is proposing a verification, validation and uncertainty quantification framework as Part 7 of ISO 23247 and is building a reference implementation of Part 5 on digital threads (NIST). | Validation of twins becomes a standard expectation, not a bespoke exercise |
| Interoperability | NIST names missing common rules as a major barrier and is developing and testing an MTConnect interface for digital twins. | More shared interfaces; confirm each against the current standard before committing |
| AI and agents | McKinsey describes a reinforcement-learning agent trained on a digital twin that scheduled four parallel lines at a metal fabrication plant, reporting "significant cost reduction and yield stability" without a percentage (McKinsey). | Twins used as safe training grounds for scheduling agents, with human approval on changes |
| Platforms | The Azure Digital Twins overview carries no retirement notice and marks 3D Scenes Studio as preview; Microsoft separates it from the digital twin builder in Fabric, also preview (Microsoft Learn). | Platform options keep shifting; confirm current availability on the vendor page before design |
| Build effort | McKinsey puts development at two to three months to over a year. | Reusable models and standards should shorten this; no source gives a figure |
The common thread is that the model matters less than the evidence behind it. Twins that carry a validation record, a named owner and a trace of the decisions taken on them will be easier to defend to auditors, plant managers and boards.
How should an enterprise start with a manufacturing digital twin?
- Pick one asset or line where an unplanned stop or a failed change is costly.
- Confirm sensor coverage, data ownership and the accuracy the decision needs.
- Build a simple twin, then validate it against real runs and record the uncertainty.
- Connect it to maintenance and planning, with a named owner for each decision.
- Only then add agents or automation, with approvals and a decision record.
To map one manufacturing decision your agents should make from twin output, and the context and policy it depends on, book a Context Workshop. See what happens in the workshop.
Frequently asked questions
What is a digital twin in manufacturing?
It is a virtual representation of a machine, line or factory kept in sync with the real one through live data. ISO 23247 gives a framework for manufacturing twins, and McKinsey calls a factory twin a real-time virtual representation of the factory.
Which standard covers digital twins in manufacturing?
ISO 23247 is the framework for manufacturing. Parts 1 (overview and general principles) and 2 (reference architecture) were published in October 2021. NIST is proposing further parts on validation and digital threads.
What are the main benefits?
Remote monitoring, earlier risk assessment before production, safe disruption testing, better team collaboration and more accurate predictive maintenance. Reported savings come from unnamed cases and should be treated as illustrations.
What is the biggest challenge?
NIST names the lack of standards for vocabulary, interoperability, trustworthiness and validation as a major barrier, and says models need validation with quantified uncertainty.
What will change in the next 3 to 5 years?
Sources point to standards for validation and digital threads, shared interfaces, AI agents trained on twins and shifting platform options. These are directions under way, not dated commitments, so confirm status on each official page.
Next Steps in Adopting Digital Twin Technology
Talk to our experts about implementing digital twin technology and smart manufacturing solutions. Learn how industries and different departments leverage AI-driven workflows and predictive decision intelligence to enhance operational efficiency. Utilize industrial AI to automate and optimize manufacturing processes, improving productivity and agility.
Ready to put AI agents to work? See an ElixirClaw demo: the agentic OS to build, orchestrate and govern AI agents across your enterprise systems.
Explore the XenonStack platform family: ElixirData · ElixirClaw · ElixirHub · NexaStack · MetaSecure · Neural AI · AgentGRC · AgentSRE · AgentAnalyst · AgentInstruct · AgentRAI · AgentScrum · AgentSearch · AgentForce · AI Aviator · SAIF Aviator · CohortIQ · XenonStack AI · XenonStack UAE