The online shop that talks to the factory
You sell online and you manufacture what you sell. On paper that is a huge advantage; in practice they are two worlds that do not talk to each other, with a person in the middle retyping. This is how that gap was closed — and what happens when AI is added to the data it produces. Names and identifying details are withheld: some of these projects were commissioned by third parties.
01 Where it started
An online shop on a standard platform, a production system living its own life, and between them someone opening both every morning to realign them by hand. Stock levels on the site were a photograph of yesterday. Delivery dates promised to customers were an educated guess by whoever knew the plant well — as long as that person was around. Every promotion had to be worked out separately, because nobody really knew what was on the line that week.
The temptation in these situations is to buy a CRM and hope it glues everything together. But a generalist CRM imposes its own way of working, and a company that manufactures has constraints no standard module knows about: batches, machine times, raw materials with their own lead times, outsourced steps.
02 What was built
A custom online shop, built around the production processes rather than beside them. No off-the-shelf CRM bent into shape: the catalogue, the availability and the delivery times come from real production data, not from a theoretical warehouse updated overnight. When an order comes in, production sees it; when production changes, the site says so.
Choosing not to use a ready-made tool was not ideological, it was about efficiency. Every step a generalist product would have solved with three screens and a nightly import is a single operation here, because the software follows the company's process instead of the other way round.
| Before | After |
|---|---|
| Stock on the site updated by hand, always behind | Availability calculated from production data, continuously |
| Delivery dates guessed by whoever knows the plant | Dates calculated from real workload and constraints |
| Web orders retyped into the management system | One order, visible to sales and to production |
| Promotions decided without knowing what is on the line | Decisions made looking at the week's production |
03 What it does today, with AI
A system that has been collecting orders and production data for years is not just software: it is a history. Models now run on that history, estimating demand by product and period and comparing it with expected production load. Not to replace the person deciding, but to give them a better question to start from: what is worth producing now, what risks sitting still, where the market is moving before the revenue figures say so.
This is the part no bought product can give you: forecasts are worth exactly as much as the data underneath them, and that data exists only because the system has been collecting it properly for years.
04 If this sounds like you
The signal is simple: someone in your company spends time keeping two screens in sync. You rarely need to rebuild everything — most projects start from the most expensive point of contact and widen from there, leaving what already works in place.
If you sell online and manufacture, or if your shop is an island compared to the rest of the company, that distance can be measured in person-hours per month. That is the first number to put on the table.
Why not use an off-the-shelf CRM or shop platform?
Because a generalist product imposes its data model and its steps, and a manufacturing company has constraints that model does not anticipate. When processes are standard, a ready-made product is the right call and we say so. When your efficiency depends precisely on how you work, custom pays for itself.
Do I have to throw away my current management system?
Almost never. In most projects the system stays where it is and the integration is built around it: reading and writing through APIs or the exports that already exist. Replacing everything is a decision to take on its own merits, not in order to add AI.
How much data do I need for forecasts to be useful?
It depends on how seasonal your market is, but the rule of thumb is at least two full cycles of whatever you want to forecast. If the data is not there yet, the first step is collecting it properly — work that has value well before any model.