Use case · Operations
AI document processing, in Thai and English.
Forms, invoices, delivery papers and applications arrive as PDFs, scans and photos, in more than one language. An autonomous system reads them, checks them against your records and prepares clean entries, while your team approves what is posted.
The problem
Why this work does not scale.
Behind most back offices sits a pile of documents that someone retypes into a system: supplier invoices, delivery notes, purchase orders, claim forms, applications. They arrive by email, LINE, portals and paper, in Thai, English or both, and no two suppliers use the same layout.
Fixed-template tools break when a layout changes, and manual entry is slow and error-prone at volume. The gap is a system that reads documents the way a careful person does, points to where each value came from, checks it against what you already know and asks when it is unsure.
How it works
From first signal to finished work.
- _01_
Collect
Gathers documents from the inbox, LINE, portals and shared folders in whatever format they arrive.
- _02_
Read
Extracts the fields that matter, including from scans and photos, and links each one to its place on the page.
- _03_
Check
Compares values with purchase orders, master data and earlier documents, and flags mismatches, duplicates and missing fields.
- _04_
Prepare
Builds the entry for your ERP, accounting or CRM system, ready for approval.
- _05_
Escalate
Sends anything unclear to a person with the document, the doubtful field and a suggested value.
Example
An example supplier invoice, step by step.
A scanned Thai tax invoice arrives by email with a PO number handwritten in the margin.
Read as supplier, date, line items and VAT, matched to PO 10482, one line differs in quantity.
✦ Fields extracted with sources, mismatch flagged, entry prepared and waiting for approval.
An illustrative case. Document types, tolerances and approval limits are set with you during mapping.
Authority
You decide how far it goes.
The same function can run at different levels. Most systems begin at Level 2, where people approve each action.
- L1 AssistExtracts the fields and shows them next to the document for a person to key in.
- L2 ApprovePrepares the full entry. A person approves each one before it is posted.
- L3 LimitsPosts routine documents that match the order and fall within your tolerances. Escalates the rest.
- L4 RunRuns intake for the document type. People review exceptions, samples and trends.
Why it works where templates fail
Template-based extraction needs a new template for every supplier layout. This system reads the document as a whole, so a changed layout or a photo taken at an angle does not stop it. It still shows its sources and says when it is unsure, so reviewers verify in seconds rather than retype.
Thai and English together
Documents in Thailand often mix Thai and English, with Thai tax invoice conventions, local date formats and company names in both scripts. Extraction is tested on your own documents before it handles live work. See our notes on designing AI for Thailand.
Governance
Guardrails built in from the first week.
Source-linked fields
Every extracted value is traceable to the page and position it came from, so a reviewer can confirm it at a glance.
Tolerance and posting limits
The system posts alone only within the tolerances and amounts you set. Everything else waits for approval.
Data protection
Scoped access, purpose limitation and a complete audit trail, designed to support PDPA compliance. See Security and our PDPA checklist.
Questions
Document processing and data entry, answered.
What kinds of documents can it read?
Invoices, purchase orders, delivery and shipping papers, forms, applications and similar business documents, as PDFs, scans and photos. We confirm the document types during mapping and test on your own samples.
Can it read Thai documents?
Yes. It reads Thai, English and mixed documents, and it flags anything it cannot read with confidence for a person.
Does it write straight into our accounting or ERP system?
It prepares the entry and, within the authority you set, posts it to the system you use. Most deployments begin with every entry approved by a person.
How is this different from OCR?
OCR turns an image into text. This system understands what the text means, checks it against your records, prepares the entry and escalates what is unclear. See autonomous AI vs chatbots vs RPA.
Further reading
Go a little deeper.
- Accounts payable and reconciliation: Read invoices, match payments, flag duplicates and prepare the close.
- AI agents for business in Thailand: a practical guide: A plain-language guide to AI agents for Thai businesses: what they do, where to start, what to check before buying and how to avoid the projects that stall.
- PDPA and AI in Thailand: a practical checklist: Using AI with customer messages, documents or HR files brings Thailand's PDPA into play. A practical checklist of what to know, decide and ask your vendor.
- Autonomous AI vs chatbots vs RPA: What each does well, where each breaks and how to choose.
Keep exploring
Related pages.
Begin
Start with this one function.
Tell Primues where work slows down. It will outline the opportunity and connect you with an architect from our team.