The term comes from industry, where a twin means a model of a physical object: a turbine, a machine tool, a workshop. In business the same word describes a model of a process: a warehouse, a queue of requests, delivery, a production line.
The difference from familiar analytics is fundamental, and it holds all the value.
A dashboard answers "what happened". A model answers "what if"
A report shows: average warehouse load 78%, peak 96%, order-picking time 14 minutes. Past time, a matter of record.
The questions the business asks sound different. What happens to picking time if the order flow doubles? Are three packers enough in December? Does a fourth line pay for itself?
The report does not answer them, because there is no such data in the past. Extrapolation lies: processes are non-linear. A 30% rise in flow may add 10% to picking time, or it may collapse the system if there is a bottleneck somewhere running near full load.
How the model is built
The process is described as a set of entities, resources, and rules.
For a warehouse: orders arrive with a known distribution across the day, each holds between one and eight items, a picker spends 40 to 90 seconds per item, there are three packers, and the shipping area holds forty boxes.
The model then runs thousands of virtual days. The output is a distribution rather than a single number: picking time averages 14 minutes, in 95% of cases it stays under 22, and under the December flow the packing queue exceeds two hours in 20% of cases.
It is the distribution that makes the model useful. Averages hide the dips, and a business breaks on the dips.
Where it works
Warehouses and logistics. Staffing across shifts, buffer-zone sizes, picking routes.
Queues and service. How many operators you need at peak, how waiting time changes with a new contact channel.
Manufacturing. Line throughput, the effect of one unit stopping on the whole flow.
Expansion planning. Whether a second workshop, a new machine, or an extra shift pays off.
When you do not need a model
Few options. If there are only two decisions and the gap between them is obvious, a spreadsheet will work it out faster.
No data on operation times. The model is built on durations and frequencies. If nobody has measured them, the first stage is measurement, and that is a separate job of several weeks.
The process changes every month. The model will be out of date before it pays for itself.
The question is about people. Motivation, work quality, and turnover are not captured by modelling.
What the client needs to provide
Three groups of data.
Process structure. Who does what, in what order, what waits for what. Collected over two or three interviews with the people who do the work by hand.
Durations. How long each operation takes, with its spread. A sample of a few hundred observations is enough, and often it already sits in the accounting system.
Incoming flow. Requests, orders, and calls by hour and day across a year. Seasonality matters more than the average.
How to check the model is not lying
Calibration is a mandatory stage. The model is run against a past period and its results are compared with fact. A gap of more than 10 to 15% on the key metrics means the model is missing something material.
Without this step you get a pretty picture with arbitrary numbers. Demand a calibration report before anyone starts making decisions on the model.
What it costs
A model of one process starts at ₽300,000. Most of the hours go on collecting the structure and on calibration; the implementation itself takes less than half.
Payback is measured through the price of the decision the model helps you make. If it is about hiring two permanent staff or buying equipment worth several million, the cost of the model is a few percent of the sum at risk. If it is about moving two desks around, skip the model and try moving them.