Generative AI (artificial intelligence) can read the documents that slow manufacturers down, such as certificates of analysis, delivery notes, supplier offers and technical specifications, and turn them into checked data in the ERP (enterprise resource planning) system. Data Department builds these workflows for manufacturers with business rules, human review for uncertain fields and models hosted on-premises or in the EU (European Union) when data protection requires it.
Which manufacturing documents are worth automating?
The best candidates are documents that arrive often, in many layouts, and whose content someone currently retypes or checks by hand:
- Certificates of analysis (CoA). Lab values per batch from raw-material suppliers, checked against specification limits before release.
- Delivery notes. Quantities, batch numbers and item codes, matched against the purchase order and goods receipt.
- Supplier offers and quotations. Prices, lead times, minimum order quantities and payment terms, compared across suppliers.
- Technical specifications and datasheets. Material properties, tolerances and standards, extracted to compare alternatives or build a searchable library.
Documents that already arrive as structured data, such as electronic delivery notes in a standard XML (extensible markup language) format, do not need AI. Read them directly. AI earns its place on PDFs, scans, photos and emails where the layout varies by supplier.
How does a document processing workflow work?
A production workflow has six steps. The model is only one of them.
- Ingest. Collect documents from email inboxes, supplier portals, scanners or shared folders.
- Classify. Identify the document type and supplier, so the right fields and rules apply.
- Extract. Use OCR (optical character recognition) where needed, then an LLM (large language model) or a vision-language model to read the fields into a fixed schema: item code, batch, quantity, each lab parameter and its value and unit.
- Validate. Apply business rules: does the quantity match the purchase order, is each lab value inside the specification limit, do units convert correctly, does the batch number already exist?
- Review. Send only uncertain or failing fields to a person, with the source document highlighted next to the extracted value.
- Post. Write approved data to the ERP, quality system or purchasing tool, and log every decision for audit.
This design keeps the model's job narrow: read accurately. Decisions about whether a batch is acceptable stay with rules and people.
When should you use rules and when a model?
Most good systems combine both. A simple guide:
| Task | Rules | Model | Why |
|---|---|---|---|
| Reading a field from a varying layout | Yes | Layouts differ by supplier; rules break | |
| Checking a value against a specification limit | Yes | Deterministic and auditable | |
| Matching a delivery note to a purchase order | Yes | Helps with fuzzy item names | Exact keys first, model for ambiguous text |
| Unit conversion and format checks | Yes | No judgement needed | |
| Summarising differences between supplier offers | Yes | Language task, reviewed by a buyer | |
| Approving a batch for production | Yes, plus a person | Quality decision must stay accountable |
If a document always comes from the same supplier in the same template, a template-based extractor may be enough. Use generative AI when variety makes templates unmanageable.
What accuracy should you target?
Set targets per field, not per document, and compare them with the current manual process. Manual retyping also has an error rate, and few companies measure it. Measure it first.
Three numbers matter:
- Field-level accuracy. Share of extracted fields that are correct, measured on a labelled test set of real documents.
- Straight-through rate. Share of documents that pass every rule and confidence check and post without human review.
- Error escape rate. Share of wrong values that reach the ERP despite rules and review. This is the one that carries business risk.
An illustrative, hypothetical example: a plant receives 2,000 certificates of analysis a month. In the first phase the team might aim for field accuracy at or above the measured manual baseline, a straight-through rate of 50 to 60%, and an error escape rate lower than today's. As rules and prompts improve, the straight-through rate rises. These figures are for illustration only; real targets come from your baseline and risk tolerance.
Where does human-in-the-loop review fit?
Human-in-the-loop means a person checks the cases the system is unsure about. The system assigns a confidence to each field and routes low-confidence or rule-failing fields to a review screen. Reviewers see the extracted value beside the highlighted area of the source document and confirm or correct it. Their corrections become new labelled examples for testing and improvement.
Review effort should fall over time. If it does not, the extraction schema or the rules need work, not more reviewers.
How do you measure the business result?
Agree the business metric before the pilot and measure the current state:
- Hours spent per month on retyping and checking documents.
- Time from document arrival to goods receipt or batch release.
- Errors found later in production, invoicing or audits.
- Supplier offers compared per buying decision, and time to compare them.
Model metrics show whether the system works. Business metrics show whether it was worth building. Our guide on how to choose your first AI project explains how to set this up.
How do GDPR and KVKK affect the design?
Manufacturing documents are mostly technical, but they still carry personal data: names and signatures of lab staff, drivers and supplier contacts. GDPR (the EU General Data Protection Regulation) and KVKK (Kişisel Verilerin Korunması Kanunu, Türkiye's Personal Data Protection Law) both apply to that data, and KVKK sets specific conditions for transferring personal data abroad.
Practical options, from most to least contained:
- On-premises open-weight models. Several capable open-weight model families can run on your own servers. Data never leaves the plant network.
- Models in your own cloud account in an EU region. Major cloud providers offer managed model services in EU regions, under your account and contracts.
- Public model APIs (application programming interfaces) with data controls. Simplest to start, but check where data is processed, whether it is retained and what the contract allows.
You can also mask personal fields before extraction when they are not needed. The right choice depends on your data, your contracts and your legal team's view. We deliver on your infrastructure: AWS (Amazon Web Services), Google Cloud, Azure or on-premises. More detail is on our generative AI assistants and agents page.
Frequently asked questions
Can generative AI read scanned or photographed documents?
Yes, with OCR or a vision-language model, but image quality matters. Low-resolution or handwritten documents need more review, so include them in the test set from the start.
Will the model approve batches automatically?
It should not. The model reads values; rules compare them with specification limits; a person remains accountable for release decisions.
How many documents do we need to start?
A few hundred real documents across your main suppliers is enough to build a labelled test set and measure accuracy. Generative models do not need thousands of examples to start working.
Can this run without sending data outside Türkiye or the EU?
Yes. Open-weight models on-premises, or managed models in a cloud region you choose, keep data inside your chosen boundary.