You are already making these decisions. Just without seeing the result first.
A digital twin is a live model of something real — a plant, a network, a building — that shows what happened, what is happening, and what could happen next. The value is not the model. It is that a decision can be tested before it is lived with.
Stop integrating systems to each other. Anchor them to the thing itself.
Connect every system to every other system that needs to talk to it. Ten systems is forty-five possible connections. Each one is a project, each one breaks when either end changes, and the integration is never finished — you are always one acquisition or one upgrade away from starting again.
Your ERP, historian, BMS, drawings and sensors disagree about almost everything — except which physical thing they are describing. Anchor each of them to the asset and the join comes for free. The tenth system costs what the second cost, and the model survives replacing any vendor in it.
That is what makes a single pane of glass possible without it being yet another dashboard: it is not a summary of other systems, it is the place they all meet.
AI needs a sandbox of your reality to reach the outcome faster.
That sandbox is a digital twin. Language models learned from a corpus of text that already existed. There is no equivalent corpus for your plant, your network or your building — so an AI that has to act in your physical world needs somewhere to practise on it.
Software AI could be trained on the internet because the internet was already written down. Physical AI has no such luxury: the situations that matter — the failure at peak load, the reconfiguration nobody has tried, the interaction between two systems that were never designed to meet — have mostly never happened, and you would not want them to happen for the first time in production.
So the data has to be manufactured. A grounded model of your reality can generate the situations that were never observed, run them thousands of times, and let an agent learn the consequences before any of it reaches a real asset. This is the same reason we found that network handover cannot be optimised from logs alone: the historical data holds one setting per cell relation, so the alternative was never recorded. It has to be generated.
That is what a twin is for, once AI is in the picture. Not a visualisation — an environment. Reality captured accurately enough that acting inside it teaches you something true about acting outside it.
AI that acts on matter — robots, vehicles, plant control, autonomous operations.
What it needs
Somewhere to fail safely. A physical agent that learns only in production is an agent whose tuition fee is paid in equipment and downtime.
Models that hold a representation of an environment and can predict how it evolves.
What it needs
A specific world, not a generic one. A world model trained on everywhere is not a model of your refinery. Grounding is what makes the prediction yours.
A live, grounded model of a real asset, kept current from its own data.
What it needs
Nothing further — this is the sandbox. The twin has been the answer for a decade; the reason it now matters more is that something has finally arrived that needs to practise.
The label moves — industrial metaverse, spatial web, Web 4.0, holographic society — and each wave renames the same capability. The European Commission has put Web 4.0 and virtual worlds into policy language, which matters mainly because it signals where regulation and funding are pointed. We would not build a strategy on the name. The capability underneath it is stable: an accurate, queryable, shared model of the physical world that people, systems and AI can all reason over. That is worth building whatever it ends up being called.
25 reasons, grouped by what they change.
Each one names what it replaces. Nothing here quotes a number we cannot source — where a figure would help, we point at the engagement instead.
See the result before you commit
The twin's first job is to let you be wrong cheaply. Every decision below is one you currently make once, in reality, with the invoice attached.
- 01
Rehearse the change before it costs anything
Instead of
Change it in production on a quiet Sunday and hope.The expensive part of a bad decision is rarely the decision — it is the six months of operating with it. Rehearsal moves the discovery to before the spend.
- 02
Compare options against the same reality
Instead of
Three vendors, three models, three sets of assumptions you cannot reconcile.When every option runs against one model of your plant, the comparison is about the option rather than about whose spreadsheet you trust.
- 03
Kill a bad idea in a week
Instead of
A pilot that runs a year because cancelling it is politically expensive.Cheap disproof is the highest-return capability a leadership team can buy. Most of the value is in the projects you never start.
- 04
Put a number on waiting
Instead of
"We'll look at it next budget cycle."Doing nothing is a decision with a cost. A twin makes that cost visible, so the do-nothing option competes honestly against the others.
- 05
Test the case you cannot test for real
Instead of
Never finding out how the system behaves at the edge until it is there.Storm load, peak demand, a failed unit at the worst hour. You cannot stage those in production. You can stage them in a twin.
Drill down, and hold people to it
This is the part most twin pitches skip. A twin is not only a picture of the asset — it is a record of who decided what, on what evidence, and what happened next.
- 06
Drill from the number to the decision that caused it
Instead of
A red KPI on a dashboard and a meeting to work out why.You click from the result to the operating change, to the person who made it, to the reasoning at the time. The question stops being "who is responsible" and becomes "what did we learn".
- 07
Every decision carries its evidence
Instead of
A recommendation in a slide deck whose underlying analysis left with the consultant.The reasoning is attached to the decision permanently. When someone asks in two years why the line was configured this way, the answer exists.
- 08
See what the organisation is about to do
Instead of
Finding out about a change when it appears in the month-end variance.Planned changes are visible before execution, not after. That is the difference between governing and being informed.
- 09
One version of what happened
Instead of
Two departments arriving at a meeting with contradictory spreadsheets.Most executive time is spent reconciling accounts of reality rather than acting on it. A shared model removes the argument and leaves the decision.
- 10
Hold vendors to their own numbers
Instead of
Accepting a supplier's performance claim because you have no way to test it.Run their claim against your reality before you sign, and against your data after. It changes what gets promised in the room.
- 11
Show the board what the plant sees
Instead of
A board pack abstracted so far from operations that nobody can question it.The same model serves the control room and the board — at different depth, from one source. Nobody is briefing from a different reality.
Reality as the integrator
The traditional answer to fragmented systems is to integrate them to each other — a project that grows with the square of the number of systems, and is never finished. The alternative is to let the physical thing be the common key.
- 12
Reality is the join everything already shares
Instead of
Point-to-point integrations between every pair of systems that need to talk.Your ERP, historian, BMS and drawings disagree about almost everything — except which physical thing they describe. Anchor on the asset and the join is free.
- 13
Add a system without re-integrating the estate
Instead of
Every new tool triggering another integration project.New sources attach to the asset rather than to each other. The cost of the tenth system is the same as the cost of the second.
- 14
Ask across silos without a data project first
Instead of
"That question would need a data warehouse initiative."Questions that span maintenance, production and finance become answerable in the room, because they all resolve to the same object.
- 15
One pane of glass that is not another dashboard
Instead of
A twelfth dashboard summarising the previous eleven.A dashboard reports the past. A twin holds the current state, the history and the possible futures of the same asset — so people stop tab-switching to assemble a picture in their heads.
- 16
Survive replacing any vendor in it
Instead of
Your operating reality trapped in a supplier's proprietary format.If the model is anchored on open geometry and open standards, swapping the platform costs a migration rather than the loss of a decade of context.
- 17
Machines and AI can reason over it too
Instead of
Knowledge that only exists in the format a human can read.The same structure that lets people agree lets agents act — grounded in your reality rather than generating plausible answers about it.
Speed, and where the money goes
Twin programmes usually fail on economics, not technology: every site starts from scratch, so cost never falls.
- 18
First value in weeks, not roadmap-years
Instead of
Feasibility study, then business case, then pilot — before one decision improves.Staged delivery means the programme pays for its next phase. It also means you find out early if it is not working.
- 19
The second asset costs a fraction of the first
Instead of
Each building or plant re-tendered as a bespoke project.The expensive part is the method, not the model. Once the pipeline exists, marginal cost collapses — this is the entire economic case for doing it at all.
- 20
Capture once, use for everything
Instead of
Separate surveys for design, insurance, maintenance and training.One capture serves every downstream need. Most organisations are paying three times for the same reality.
- 21
Spend on the constraint, not the shopping list
Instead of
Buying the sensor package the vendor recommends.A scored assessment tells you which gap actually blocks the decision you care about. We have told clients to buy less than they planned.
- 22
Use the data you already have
Instead of
"We need to fix our data before we can do anything with AI."Most organisations are ready far earlier than they are told. The constraint is usually sequencing, not data quality.
Risk, compliance and the knowledge that walks out
The quiet benefits — the ones that only become obvious the day you need them.
- 23
Prove what you did, to a regulator
Instead of
Reconstructing a decision trail from email after the request arrives.Time-stamped, evidence-linked records of state and change. Under NIS2 and the EU AI Act this stops being good practice and starts being the audit.
- 24
Find the seam before an attacker does
Instead of
An asset register that was accurate the year it was made.You cannot defend what you cannot enumerate. A live model of what exists and how it connects is both the security baseline and the compliance artefact.
- 25
Keep the knowledge when the expert retires
Instead of
Thirty years of judgement leaving with one person.The reasoning behind how this asset is actually run gets captured where it is used, not in a handover document nobody opens. For most industrial businesses this is the largest undeclared risk on the register.
Four things we can point at, and four we won't.
Every entry below was read at its primary source, and says whether it is a measured result or a projection. Those two get blended constantly in this industry.
- Measured resultJune 2025
Collision checks in production planning cut from roughly four weeks to three days.
BMW Group, in its own press release
The clearest published before/after we could verify at source. BMW separately states digital twins are in use across 30+ production sites.
Source ↗ - Measured result2023
A plant reached virtual start of production roughly two years before real operations began.
BMW Group, on its Debrecen plant
The whole factory was commissioned in simulation before it existed physically — the clearest public example of rehearsing reality at scale.
Source ↗ - FrameworkCurrent framework
Autonomous Networks maturity is defined on a six-level scale, L0 to L5 — the same shape as automotive autonomy.
TM Forum
Useful because it gives telecom operators a shared vocabulary for where they actually are. TM Forum's Catalyst projects now pair digital twins with agentic AI specifically to reach the higher levels.
Source ↗ - Policy document11 July 2023
Web 4.0 and virtual worlds are the subject of a formal European Commission strategy.
European Commission, COM(2023) 442 final
Worth knowing mainly because it signals where EU regulation and funding are pointed. We would not build a strategy on the label.
Source ↗
These four circulate constantly in digital-twin marketing. We checked them, and each one fails for a different reason. If a supplier quotes you one of them, ask where it came from.
"65% less downtime and $52M in annual savings" at a named consumer-goods manufacturer
Quoted everywhere, always crediting the same 2019 vendor article. We fetched that article: it contains none of those numbers. The chain is aggregator-to-aggregator with no traceable primary source.
"75% of organisations implementing IoT already use digital twins"
The underlying survey found 13% actually using them and 62% planning to. The 75% adds the intenders to the doers. It is a forecast wearing the clothes of a measurement.
"The digital twin market will reach $X billion by 20XX"
A single search returns four mutually contradictory forecasts. None of them tells you whether a twin is right for your plant.
"The Centre for Digital Built Britain is building the UK National Digital Twin"
Out of date since September 2022 — that programme moved into government. Still repeated in a great deal of current marketing.
We went looking for an independent, non-vendor study measuring return across a population of digital-twin deployments. There isn't one. Every credible measured result in this field is a single organisation reporting on itself — useful, but not a benchmark. So treat any confident industry-wide ROI number, including one pointed at you by us, as a forecast until someone shows you the method behind it. That is also why our own case studies name what could not be validated.
None of this requires a transformation programme.
Pick the asset where you make the most expensive recurring decision with the least confidence. Twin that one. The method is what carries to the next.
Prefer to check your own position first? The Twin Score takes five minutes, scores you against the SMILE model, and there is no email wall.