Reducing manual work in bike leasing

The company wanted to accept external dealer quotations as a new ordering channel. That changed where order information entered the process: instead of starting with a structured internal form, the workflow had to begin with a quotation supplied by a dealer.

Manual re-entry would have shifted too much work to customers or employees. The quotations also varied considerably, so the challenge was not just reading text from a PDF. The workflow had to understand which amounts and line items represented the bike, accessories, upgrades, insurance, or items that could not be leased.

Dealer quotations arrived as varied PDFs from across Belgium. Layouts, terminology, line items, fees, discounts, VAT handling, and the way prices were presented could differ from one quotation to the next.

The workflow needed to preserve an important distinction. Catalogue prices were needed for residual-value calculations, while quotation prices could include discounts and VAT differences. Treating every number as interchangeable would create incorrect orders even when the extraction itself looked plausible.

I proposed the automation and owned it end to end, from business and technical analysis through ideation, proposal, proof of concept, backlog, development, integration, and iteration after launch.

The first production version used Google Document AI to turn a quotation into structured data. The application extracted the needed values, applied business checks, attempted order creation, and reported issues for employees to resolve. Employees could inspect the quotation and adjust a draft order rather than asking a customer or dealer to re-enter information that was already present in the document.

The first production version used a practical pricing heuristic: the largest amount was assumed to be the bike. That was a useful starting point, but it did not reliably distinguish accessories and upgrades.

Later iterations improved document interpretation and classification as real quotation variation and production exceptions became clearer. The workflow was developed with the people who operated it, using the cases that needed attention to decide what the next version should handle better.

The workflow is used across thousands of orders. More than nine in ten orders are acceptable without further correction, while employees still have a clear place to inspect and resolve exceptions.

Customers or dealers only need to be contacted when the quotation itself is wrong. That keeps human attention focused on the cases that need judgement instead of making people repeat information that the quotation already contains.

AI automation is rarely a one-shot implementation. Build the process closely with the people who operate it, test against a broad enough set of real inputs, and let production exceptions guide the next version.

In this case, the useful boundary was clear: document interpretation could reduce manual work, but pricing and item checks and employee review still needed to protect the order process.