Digital twins

Digital twins for business, explained

A digital twin is a virtual model of a real system, continuously informed by data, that can be used to understand how the system behaves and to test decisions before making them in the real world. Engineers use digital twins of jet engines and power plants; medical researchers use them to model how a body responds to treatment. These are domains where uncertainty is high and the tolerance for error is very low.

Businesses share the first property. Most would like to share the second.

Why a business needs a “world model”

Business outcomes are produced by many internal and external factors acting at the same time. Understanding which drivers really matter — and how much to invest in one initiative versus another — too often becomes a half-data, half-art exercise in which politics biases the decision.

CFOs sit at the centre of these decisions but rarely have a consistent model of how the enterprise works and how turning each “knob” changes results. They rely on department-level analyses that are deep in their own area but lack the full picture — and whose impacts never add up.

What a commercial digital twin is

A commercial digital twin treats the company as a dynamic system: many parts interacting in complex, often partly unknown ways. It models the entire commercial engine — product, brand, PR and word of mouth, marketing, customer intent, sales and support, the customer base — with all the dependencies between them, in one consistent representation of the business.

Relationships are built as they work in reality, using both internal data (spend, sales activity, inventory, product changes) and external data (seasonality, macro conditions, competition). Lags and feedback loops are explicit: a campaign today can raise brand awareness that pays off months later; satisfied customers feed word of mouth.

Borrowed from engineering and medicine

To build it we borrow techniques already proven in industrial system analysis and medical research — control theory and Bayesian dynamic state-space models — and adapt them to a business context. Our scientists come from domains such as health and defense, and work alongside people with hands-on commercial experience in advertising, marketing, sales and inventory planning.

Uncertainty as a tool

A digital twin does not pretend to know exactly how the business works. With “small data” and many overlapping drivers, many models can fit history equally well. So instead of picking one, the infrastructure tests thousands of model variations and converges on three or four competing explanations.

• When they point in the same direction, the decision is robust: you know a move is supported whatever you believe about, say, how quickly TV works.
• When they disagree, the disagreement itself is valuable: it frames an informed decision or guides a targeted test.

Recommendations are then risk-adjusted, the way a CFO would treat any investment, rather than going all-in on the option with the highest average return.

Under the hood: three technical layers

Technically, the twin is organised in three layers, each adapted from methods matured over twenty years of biomedical research.

1. A causal map of the business

The first layer is a directed graph of internal and external drivers — demand, seasonality, price, competition, marketing and sales actions, inventory, internal events — together with their dynamics: impact delays, saturation and indirect effects. It is the direct equivalent of a disease-progression model: networks of latent stages and observable biomarkers fitted to longitudinal cohorts, where the underlying state cannot be observed and must be inferred from sparse, noisy, indirect measurements. A company’s “brand health” or “latent demand” behaves exactly like that. The equations that drive our optimiser were first written to track the progression of Alzheimer’s disease.

For each driver we set plausible ranges for coefficients, delays and saturation points, and encode them as structured Bayesian priors — so the model starts from what is physically and commercially credible, instead of letting collinear data produce implausible answers.

2. A validation engine, not a single model

The second layer tests more than 1,000 Bayesian model variants per business outcome: hierarchical regressions, Bayesian neural networks and multi-stage models, each encoding different assumptions about delays, saturation, interactions and indirect effects. The protocol borrows from bioinformatics — gene-selection methods and causal mediation analysis, used in medicine to expose dataset bias in deep-learning models. Every variant is scored on three criteria:

• Explanatory gain — how much it explains beyond a simpler baseline
• Out-of-sample predictive gain — how well it forecasts periods it has never seen
• Resilience to “natural experiments” — whether it behaves correctly around real shocks observed in the data (outages, price changes, launches, pauses)

Variants that pass are grouped into families and combined through Bayesian model averaging. The result is not one “best fit” but a calibrated distribution of plausible answers — the uncertainty is measured, not assumed away.

3. Forecasting and optimisation under uncertainty

The combined model acts as the twin itself. It forecasts, explains deviations from plan, detects inflection points and runs multi-objective optimisation of resource allocation under uncertainty: ideal allocations are computed for different investment levels and different degrees of risk aversion, and a scenario planner tests macro and business hypotheses. Bayesian neural network architectures, originally designed for echocardiographic assessment, bring deep-learning flexibility with calibrated uncertainty to this predictive layer.

In production, the pipeline runs on a cloud data warehouse, with inference orchestrated at scale so that thousands of variants can be re-estimated as new data arrives.

Where the method is going

Today, building the causal map for a new client is the longest step: a senior scientist identifies drivers, delays, interactions and coefficient ranges. Standard MMM tools such as Robyn or Meridian also require the user to specify this structure, and automatic causal-discovery methods remain fragile on highly collinear corporate data. Our current research, in partnership with INSEAD, uses AI to extract valid drivers, relationships, dynamics and coefficient ranges for each sector from the scientific literature, and to use them as structured priors — with the validation engine remaining the final arbiter.

Questions a digital twin can answer

Because it covers all key drivers, the twin is modular: new business questions plug into the same model.

• What drives my business — and what explains the latest acceleration or deceleration?
• What was the real impact of a product launch, a campaign or an investment?
• How should we split budget between marketing and sales, and how much to invest in brand?
• What happens if the macro situation worsens — and where can we safely cut costs by 10%?
• Should we expand into a new geography, and what should we expect?

One model everyone can trust

The real value is alignment. Finance, operations, marketing, sales and product work from the same interpretation of the business, expressed in finance’s language of profit and risk-adjusted returns. Conclusions that confirm current beliefs and conclusions that challenge them are presented side by side, so the executive committee can take a fully informed decision.