I took over spare-parts supply at BPK Auto in Ust-Kamenogorsk in the mid-2000s, at the age of 25, moving from selling cars. The company was the general dealer for VAZ in Kazakhstan, assembled Niva models in the city, and was taking on Skoda, Chevrolet, and Kia while building service centers across the country to meet the expectations of those brands for a dealer. My job was to keep parts flowing to all of it. Several thousand vehicles came through in a month, and a single car contains something like 50,000 separate part numbers.
The service teams wanted as many of those as possible sitting on the shelves. From where they stood, that made sense, because they were judged on whether a mechanic could start a repair the same day, whether a customer complained, and an ISO audit was running over the service operation at the same time.
From the procurement side, the arithmetic looked different. Some parts cost two or three thousand dollars and were fitted once or twice a year, and holding them across four model lines tied up a large share of the company’s working capital. I built the forecast in spreadsheets and checked it against actual sales once the month closed in 1C. To run it, I needed the service side’s own records: how many vehicles of each type they had seen, and what had been repaired. I asked for them for months. They would not release them.
Some of that was habit; a warehouse full of everything being what a well-run operation had looked like for the previous 40 years. More of it sat in how the two departments were counted. My result was recorded the moment the parts moved into the service center. Theirs was recorded much later, when the car left the workshop. Nobody was measured by the number that would have made the forecast work.
The forecasting was the straightforward part of that job. The difficult part was building an operation in which people shared the information behind the forecast and were expected to act on its findings.
Why a Better Model Would Not Have Changed the Outcome
Put one of today’s AI demand-forecasting systems into that same warehouse and very little changes. The forecast would be far better than anything I produced by hand. It would still depend on service records that nobody was willing to release, and where records did exist, it would be reading entries written up at the end of a long shift. It could not grant the purchasing manager authority to reorder against it, nor could it alter the measures by which the service department was judged. Better intelligence does not repair the path between operational data and operational action.
Every party you add to that path lengthens it. A distribution network across Kazakhstan and Kyrgyzstan can run from a manufacturer to over 45,000 small retail outlets, served through more than 18 distribution centers, with 1,391 field staff taking and checking orders on the ground. An order begins as a conversation in a shop. It then passes through a sales representative, a warehouse, a driver, a distributor’s system, and eventually manufacturers, and each of them records a slightly different fact about it at a slightly different moment. Where retail is this fragmented, several parts of the business end up planning against their own version of the same week.
What It Took to Close That Gap
The system I keep coming back to is the mobile application we built at TezCo Trade and put in the hands of the field sales force, because it dealt with the problem I had failed to solve at BPK Auto.
Before, a representative visited a shop, agreed on an order, and wrote it down. The distribution center planned its stock based on what had been ordered last time, and the manufacturer saw the market through a summary assembled weeks later. Nobody in that chain was withholding anything deliberately. The information simply had no route from the shop counter to the people planning against it. The application recorded the order, the stock already on the shelf, and the outlet’s state at the time of the visit, and delivered all of it to the warehouse and the planning team the same day.
The technical part was the smaller half of the work. Representatives were now producing a record that made their own working day visible, which was close to the position the mechanics had been in years earlier.
What produced that gain was a single operating record shared by teams that had until then worked from different versions of the same supply chain. The same exercise was repeated for warehouse management, transport, the EDI links with manufacturers, and the analytics we now supply back to our partners. Each time, it was a piece of software first, and then the slower question of whose job it changed.
Three Things a Model Cannot Supply
The two experiences point to the same three conditions, and none of them arrive with the technology.
The first is reliable data. A model cannot create records that people do not collect, and it cannot correct careless entries because nobody is held accountable for their accuracy.
The second is decision authority. An accurate forecast is worth nothing if the person who receives it cannot change the order, stock level, route, or supplier.
The third is aligned incentives. A warehouse judged only on never running out will keep holding more than it needs, no matter what the forecast says. People who ignore a better answer are usually responding to what their organization still rewards them for, which is the old decision.
The Perfect-Answer Test
Before spending on a model, put the operation through a simple check. Imagine that the perfect answer to the decision you most want improved is already sitting on the right person’s desk tomorrow morning. Then work through three things: whether you have the data to produce that answer regularly rather than once, whether the person receiving it can change the decision, and what would make them set it aside.
Run BPK Auto through it, and the answer arrives quickly. Say a system had produced an exact demand figure for every part in the catalog. The service records behind that figure were held by people who would not release them, so it could not have been produced twice. The purchasing decision pointed to a department measured on shelf availability, so nothing in the way of the service teams was judged that would have moved. The figure would have been correct, and it would have sat there.
If a perfect forecast would change nothing, the problem sits upstream of the model, in the data flow, the decision rights, and the incentives.
What Readiness Actually Means
Central Asia does not need less ambition around AI. It needs a more operational definition of readiness. The companies that get something back will be the ones that made their data reliable first, gave the person receiving the answer the authority to act on it, and stopped paying their teams for the decision they are trying to replace.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of the publication, its affiliates, or any other organizations mentioned.
