Before you build it, simulate it: the case for a digital twin of everything.

Most decisions inside most organisations still get made the same way. Someone forms a view, builds the thing, ships it, and finds out whether it worked by watching what actually happens. If it goes badly, the response is damage control after the fact, not a course correction before the fact. Researchers who study how large platforms make decisions have a blunter name for this pattern, launch-and-hope. It's not that people aren't trying to be careful. It's that most organisations genuinely have no way to look around the corner before they commit.

We've been sitting with a question that we don't think has a fully worked-out answer yet, but that we suspect points somewhere real. What if that corner could be looked around first? Not through more research, more surveys, more historical data pulled from what already happened, but through an actual working simulation, a digital twin, of the environment a decision is about to land in, tested and stress-tested before anyone commits to the real thing. We think this applies to a new product feature. We think it applies to a government policy. We increasingly suspect it applies to almost anything with people on the other end of it.

01Why most simulations get this wrong

The obvious objection to simulating human behaviour is that people are hard to predict, and that's true, but it's worth being precise about why. Simulation researchers who work on this problem draw a distinction between two kinds of data, and it's a distinction we think most conversations about digital twins skip past too quickly. There's the what, what people actually do, their purchases, their clicks, their choices, measured directly with no interpretation required. And there's the why, the motivations, context, habits and constraints sitting underneath that behaviour, the part you can only get at by asking people, or by understanding the mechanism well enough to infer it.

Neither one alone is enough. A model trained purely on historical behaviour can tell you what happened before, but it has no way of knowing whether that pattern still holds once the underlying conditions change, and conditions changing is exactly the situation you're usually trying to simulate for. A model built purely on what people say about themselves runs into a different problem, people are not reliable narrators of their own behaviour. We rationalise after the fact, answer aspirationally, and often simply don't know why we did something. The gap between what people say and what they actually do is itself a well-documented, patterned thing, not just noise to average out.

A genuinely useful digital twin of a market, a workforce, or a population has to hold both the what and the why together, calibrated against each other, so it can be asked questions about situations that haven't happened yet rather than just questions about situations that already have.

02Why language models specifically are good at the why

There's a reason we think this idea has become newly tractable rather than staying purely theoretical, and it comes down to what large language models were actually trained on. An LLM isn't trained on a narrow slice of behavioural data the way a pricing model or a traffic model is. It's trained on an extraordinary cross-section of humanity's own recorded expression, what people have written, argued, confessed, explained, and reasoned through, across every background, era and circumstance that left a written trace. That gives it an unusual and, we think, underused capacity, the ability to distil a persona on demand.

Tell a model "now you are a starving artist, respond to this," or "now you are an Asian parent, react to this," or "now you are a Chinese immigrant from a specific place and era, decide what you'd do here," and it can take on that persona with a texture that goes well beyond a demographic label. Not perfectly, and not without real limitations worth being careful about, but with enough fidelity to the motivations, constraints and lived context that come bundled with a background like that, which is exactly the why layer we described above, made addressable on demand rather than something you have to painstakingly hand-code into a traditional agent-based model rule by rule.

03This is already happening at national scale, in Singapore

Here's the part that moved this from a hypothetical to something we think is worth writing about now rather than in five years. Singapore has already built a working version of this, not as a single finished product, but as a growing ecosystem of national-scale simulation infrastructure, some of it running today.

Virtual Singapore, the country's 3D digital twin, provides the geospatial base layer, buildings, terrain, infrastructure. But the part that actually matters for the argument we're making sits in the layer built on top of it. A*STAR's Institute of High Performance Computing now describes itself as the foundational engine for Singapore's digital twin ecosystem, and its stated approach is explicit about combining physics-based precision, data-driven AI, and what its own researchers call "the nuances of human behaviour."

Two live programmes make this concrete. The Singapore Integrated Transport and Energy Model blends simulation of the national power grid with real driving behaviour and vehicle travel patterns, used specifically to plan long-term EV charging infrastructure, a genuine agent-based model of how people actually move, feeding directly into national infrastructure decisions. The Centre for Energy and Emissions Modelling, described by A*STAR as a first-of-its-kind integrated digital twin, simulates how decarbonisation efforts across energy, industry and transport sectors interact with each other before those policies are committed to. A*STAR's own framing of why this matters echoes the Zillow lesson almost exactly, digital twins are needed most, in their words, "when the experimental means of validating human biomedical research are unsafe, too expensive, too time-consuming, or involve vulnerable populations." Read policy for biomedical research and the same sentence holds.

It's also worth being honest that not every attempt at this has landed cleanly elsewhere. Argentina's Ministry of Human Capital launched a comparable Digital Twin initiative in 2026, intended to simulate the impact of social policy, poverty programmes, subsidies, before rollout. The idea itself sits in the same family as what Singapore is doing. But the project also drew genuine public criticism, over its ties to outside data and analytics vendors, and over unanswered questions about the legal basis for using citizens' data this way. We think that's a useful caution rather than a reason to abandon the idea, the technique is sound, but who builds it, what data it runs on, and how transparently it's governed clearly matter as much as whether it works technically.

"A digital twin is not just a model of a place. It's a rehearsal space for a decision, and like any rehearsal space, it only earns trust if you can see how it was built."

04What if this scaled down, not just up

Everything above operates at national scale, transport networks, power grids, entire sectors. Our actual question is smaller and, we think, more immediately useful. What if this same rehearsal space existed for a single business, a single service, a single app, before it went anywhere near a real customer?

Picture a business preparing to change its pricing, launch a new product line, or restructure a customer-facing process. Today, that decision usually gets made on a mix of market research, competitor benchmarking, and a healthy amount of professional judgement, then it ships, and the business finds out whether it worked by watching real customers respond in real time, with real consequences if it doesn't land. A working digital twin of that business's own customer base and operating environment, built the same way Singapore builds its national models, calibrated on real behavioural data and real context about why customers behave the way they do, would let that same decision get tested first. Not once. Repeatedly, cheaply, in private, the same way A*STAR tests decarbonisation policy before it touches the real grid.

The same logic extends to internal decisions too, a workforce policy, a new process, a restructuring. Simulating how a team is likely to actually respond, grounded in real behavioural patterns rather than an HR consultant's best guess, is a version of exactly the same idea, just aimed inward instead of outward.

05We've actually tried a version of this ourselves

We said earlier that this article isn't meant to showcase a finished solution, and that's still true. But we'd rather be upfront than coy about something relevant, we've already experimented with the concepts we've just described, not as a thought exercise, but as a real, running simulation.

Last year, at our Hack to the Future Thinkathon with Temasek Polytechnic's School of Informatics and IT, we built our own simulation and used it to test how an insurance agency could be run entirely by AI agents, inside a realistic economic market context, compressed across a simulated 20-year horizon. It wasn't a scripted demo. Different teams' agents, working from different governance parameters, ended up in genuinely different places, one configuration grew capital from roughly SGD 500,000 to over SGD 304 million, another configuration, built on the same underlying model but different guardrails, went bankrupt. That divergence was the entire point. It's exactly the kind of consequence you want a system to experience before it ever touches a real market, real capital, or real policyholders.

That's really the deeper argument underneath everything in this piece. We think a good simulation isn't just a nice-to-have for testing business decisions. It's close to a foundational piece of the stack for agentic AI generally, because it functions as a pretraining playground. An agent that has been allowed to make decisions, watch them play out, fail in expensive ways, and adjust, inside a simulation with no real-world consequences, is a fundamentally more trustworthy agent to eventually let loose on the real thing than one that has only ever been evaluated on a static benchmark. Experiential learning without real-world stakes is exactly what a flight simulator does for a pilot. We think agentic systems need the equivalent, and we don't think that infrastructure exists in any general, reusable form yet.

Since that first run, we've kept building on it. We've been working to give the simulation deeper dimensions, more variables that actually matter to how an economy or an organisation behaves, longer simulated time horizons so slower, more structural dynamics have room to show up, and a more complex interplay between different actors and situations acting on each other simultaneously rather than in isolation. The honest goal, and we're not there yet, is to push it toward something closer to a universal simulation substrate, one adaptable enough to model an insurance market one week and a completely different kind of business or policy question the next.

06Where Trusted Hub fits into this

The foundational pieces a business-scale digital twin needs already map closely onto the stack we've spent years building. Trusted DataVerse handles the multimodal ingestion and enrichment of an organisation's real operational data, the raw material any simulation needs to be calibrated against reality rather than assumption. Trusted KnowledgeVerse turns that enriched data into a governed, continuously updated knowledge base a simulation can reason over. Trusted ActionVerse is the agentic layer capable of modelling how a change might actually ripple through that knowledge, the same layer that ran our own Thinkathon simulation last year.

A digital twin of a business is only as trustworthy as the what and the why sitting underneath it, exactly the same principle Singapore's own national models are built on, and exactly what we tried to prove out ourselves at a smaller scale. We think that's genuinely within reach for Singapore's businesses, not next decade, but soon.

We don't have a finished product to announce here. What we have is a question we think is worth asking out loud, and a growing pile of evidence, from Zillow's very real $304 million mistake to Singapore's very real national infrastructure models, that the answer is probably yes, more decisions deserve a rehearsal first. We'd genuinely like to hear what you think this could look like for your own business, because we suspect we're still early in figuring that out ourselves.