Unilever is rolling out factory digital twins with an unusual sequencing rule: business results come first, technology second. Supply Chain Dive reported on October 17 that the company puts measurable outcomes ahead of the novelty of the model. That sounds like management language, but it has a concrete implication. A digital twin is only useful if it can be tested against a counterfactual, and most manufacturing analytics programs never build one. The difference between a twin that informs decisions and a twin that decorates a dashboard is the presence of a baseline, a comparison and a cost.
Key Conclusions
A digital twin is a simulation of a physical process that runs on live or near-live data. Its value depends entirely on the decision it changes. If the model recommends a change in batch size and nobody can measure what happened without it, the twin has produced an opinion rather than evidence.
The retail context raises the stakes. September retail sales rose 6.7 percent, largely on price, and a New York Fed analysis attributes 2.9 percentage points of consumer goods inflation to tariffs. When input costs are moving, a manufacturing model that cannot quantify a cost trade-off is not a decision tool, it is a distraction.
What a Twin Needs to Be Testable
A baseline before the model
Every twin needs a pre-period baseline measured with the same instrumentation as the post-period. Without it, any improvement can be attributed to seasonality, mix or a dozen other factors. The baseline is boring work and it is the difference between a pilot that scales and a pilot that gets quietly retired.
A counterfactual, not just a forecast
A forecast says what will happen. A counterfactual says what would have happened otherwise. Only the second supports a claim of causation, and only a claim of causation survives a finance review. Twins that produce counterfactuals get funded; twins that produce forecasts get questioned.
Best Practices
Tie each twin to one cost line
Pick one cost line per deployment: energy per unit, scrap rate, changeover time or overtime hours. A twin optimized for several objectives at once is hard to evaluate and easy to defend with anecdotes. One cost line, one baseline, one comparison window.
Instrument the floor before the model
Sensor coverage determines what the twin can see. Before deploying a model, audit which process steps have data and which are inferred. Inferred steps are where the model will be wrong, and knowing that in advance is what allows the team to state confidence rather than imply certainty.
Common Mistakes
The first mistake is buying the model before defining the decision. A twin that answers a question nobody is asking will not change behavior, no matter how accurate it is. The second mistake is measuring success with model metrics. Accuracy, latency and coverage are engineering metrics; a factory manager needs a cost or throughput number.
The third mistake is running the pilot in a stable period. A twin proves its worth when conditions move, because that is when the counterfactual matters most. Piloting during calm months produces a result that cannot be generalized, and the program stalls at the first disruption.
An Evidence Checklist for Analytics Programs
The checklist has four items. First, a named decision owner who will change a process if the model says so. Second, a baseline measured before deployment with identical instrumentation. Third, a comparison window that includes at least one demand or supply shock. Fourth, a cost line that finance already tracks, so the result does not need a new accounting definition.
Applied to retail operations rather than factories, the same checklist works for assortment, replenishment and labor scheduling. The common failure is not model quality but the absence of a decision owner. Programs with an owner and a baseline tend to survive; programs without them tend to produce dashboards that nobody opens after the pilot.
Summary
Unilever's sequencing rule is a useful correctionThat difference also determines whether the program survives its second budget cycle, when the novelty has worn off and only the numbers remain. to how analytics programs are usually sold. Technology first produces pilots; results first produces decisions. The difference is whether a baseline and a counterfactual exist before the model does.
For retail and supply chain teams, the practical test is simple: name the decision, name the owner, and name the cost line before deploying. If any of the three is missing, the twin is not ready, regardless of how good the simulation looks.
Data Sources
- Unilever factory digital twin rollout：https://www.supplychaindive.com/news/unilever-puts-business-results-first-in-factory-digital-twin-rollout/833011/
- September retail sales up 6.7%：https://www.retaildive.com/news/monthly-retail-sales-from-the-us-commerce-department/574252/
- Tariffs add 2.9 points to goods prices：https://www.retaildive.com/news/tariffs-pushed-up-consumer-goods-prices-ny-fed-inflation-elections-tax-taxes/832701/
- Gen Z and AI-driven shopping：https://www.retaildive.com/news/gen-z-gen-alpha-holiday-shopping-retail-trends/831838/
FAQ
What makes a digital twin testable?
A:A pre-period baseline measured with the same instrumentation, plus a counterfactual that shows what would have happened without the model.
Why is a forecast not enough?
A:A forecast describes what will happen; only a counterfactual supports a causal claim that survives a finance review.
How many cost lines should a twin target?
A:One per deployment. Multiple objectives make evaluation ambiguous and invite anecdotal defense.
What is the most common failure?
A:No decision owner. Without someone who will change a process, the model produces a dashboard rather than an outcome.
When should the pilot run?
A:Across a period that includes a demand or supply shock, because that is when the counterfactual matters most.
Does this apply outside manufacturing?
A:Yes. Assortment, replenishment and labor scheduling follow the same evidence logic.
