Custom AI-powered ordering system for a wholesale plant importer

A wholesale plant importer working with 18 regular suppliers across the Netherlands, Germany, Denmark and Italy processed around 3,500 SKUs a month. Every supplier communicated differently: one sent Excel files with prices and availability, another PDFs with species lists, a third texted on WhatsApp with pallet photos, a fourth ran their own B2B portal.

A 2-person purchasing team spent over 240 hours per month rekeying orders into the internal system, checking stock and calculating costs per supplier. Invoices arrived 2–6 weeks late, so margin control was always after the fact. It was hard to answer quickly: how much do we actually make on supplier X in the spring season?

Horticulture added its own complications: Latin plant names are often misspelled, the same species ships under different trade names (Strelitzia reginae = Crane Flower = Bird of Paradise), and seasonality changes availability day to day. Every new supplier meant two weeks of learning curve for a purchaser to decode their specific format.

Weeks 1–3: workshops with purchasers and discovery on a sample of 250 historical orders. We built a species dictionary (600 Latin names with trade synonyms and translations) plus a map of each supplier's commercial policies. That dictionary became the knowledge base OpenAI later consulted when classifying ambiguous items.

Build took 6 weeks, three layers. First layer — OpenAI GPT-4o as parser: identifies document format (Excel, PDF, email, WhatsApp), extracts line items (species, size, quantity, price), normalizes Latin names to the species base, flags ambiguous items for human review. Second layer — n8n as orchestrator: routes orders from suppliers (Outlook, WhatsApp Business API, FTP, custom webhooks), pipes them to the AI parser, syncs to the ordering system, generates pick lists for the warehouse. Third layer — a custom web app (Node.js + React) for purchasers: daily order pipeline, cost-per-supplier calculator with margin forecast, automatic invoice pull from supplier portals and AI parsing on top.

The last 3 weeks were a staged rollout. First the 5 easiest suppliers (Excel + standard formats), then 8 medium ones, finally 5 of the hardest (WhatsApp, custom portals, atypical naming). Each stage had a week of 100% human-verified pilot, then a transition to auto-mode with verification only on flagged items. The auto/manual ratio was set adaptively — the model learns each supplier's policies.

After the full rollout, the 2-person purchasing team moved to higher-value tasks: sourcing new suppliers in Spain and Italy (previously no time for that), negotiating commercial terms with existing ones, monitoring delivery quality. The supplier count grew from 18 to 27 in the first year post-rollout — with no headcount increase.

The commercial director now sees a real-time margin-per-supplier dashboard in the app, with weekly forecasts based on incoming orders. „Is it worth buying from supplier X this week?” is now an on-the-spot decision, not an after-the-fact one. In spring 2026 the company increased trade volume by 28% year-over-year with the same back-office headcount.

An unexpected effect: Latin name recognition with synonyms surfaced two cases where the same species was being bought from two different suppliers under different names at prices differing by 18%. The annual savings from that observation alone covered 40% of the rollout cost.