Reading tech packs with AI: what works and what does not
Lessons from teaching a model to read buyer tech packs, from strikethrough trims to size runs split across pages.
A buyer's tech pack is the most important document in a garment order, and one of the hardest for software to read. It can be a 40-page PDF, a set of scanned sheets or photos of printed pages, laid out differently by every brand.
While building the AI tech pack reader in TB-ERP, our garment ERP, we learned a lot about where language models help and where they need guard rails.
What works well
- Pulling structure out of messy layouts. Modern models are good at finding the BOM table, colourway matrix and size run even when every brand formats them differently.
- Treating many files as one document. Tech packs often arrive as several PDFs and images. Reading them together lets the model connect a BOM on one page with colourways on another.
- Keeping long descriptions intact. Fabric descriptions can run to a paragraph. The model can keep the full text and still pick out GSM, construction and yarn count into separate fields.
What needs guard rails
- Strikethrough and cancelled items. Brands cross out dropped trims rather than deleting them. The model must be told explicitly to skip anything crossed out or marked cancelled.
- The urge to fill gaps. Left unchecked, a model will guess a missing value. Our prompts require it to extract only what is written and leave anything else empty.
- Which code is the style number. A tech pack may show a style code, an article number and a model name. We give the model a strict order of priority.
- Supplier names split across columns. Name and supplier ID often sit in separate cells and need to be combined consistently.
Checking, not trusting
We never send extracted data straight into costing. TB-ERP compares the tech pack with the buyer's purchase order, style by style and size by size, and flags any difference for a merchandiser to resolve. That cross-check, rather than the extraction alone, is what makes the result safe to use. In production, AI cross-checking in TB-ERP reaches 99% accuracy.
What changed for merchandisers
Instead of retyping each tech pack, merchandisers now review what the AI extracted and fix the few flagged items. Combined with CAD and IE imports and bulk costing, TB-ERP users report around 80% less time spent on costing, bulk orders, tracking and management.
Key takeaways
- Models handle varied tech pack layouts well when asked for a strict schema.
- Tell the model what to ignore: cancelled, crossed-out and missing values.
- Cross-check against the PO before anything reaches costing.
Have a project in mind?
Tell us what you want to build, improve or automate. We reply within one working day.