Why This Document AI Update Matters to Your Business
If your team still retypes information from invoices, purchase orders, delivery notes, contracts or customer forms, a new document AI release deserves your attention. LandingAI has launched Agentic Document Extraction Gen2, introducing a document-processing system designed to understand structure, identify important fields and show exactly where extracted information came from.
For a Malaysian SME, this is more than an OCR upgrade. It points towards a practical shift: software can process documents while giving your staff a verifiable trail back to the original page, line or word. That matters when you handle multilingual paperwork, scanned forms, handwritten notes, tables and documents that must be checked before entering your accounting, inventory or customer systems.
LandingAI says Agentic Document Extraction Gen2 is generally available and can be accessed through its playground, cloud environments, private cloud deployments and on-premises installations. Source: Marktechpost
What Happened
LandingAI rebuilt its document intelligence platform around two parsing models: DPT-3 Pro and DPT-3 Verity. The company describes Gen1 as treating a document like a flat list of text chunks. Gen2 instead represents the document as a hierarchy containing the document, pages and individual blocks such as text, tables, figures, signatures, logos and barcodes. Source: Marktechpost
DPT-3 Verity is intended for digitally created documents, high-volume text, tables and straightforward form fields. It returns a bounding box and confidence score for every word. DPT-3 Pro first analyses page layout, then recognises different blocks and reading order. It is designed for scanned documents, handwriting, non-Latin scripts, mathematical notation, figures and more complex layouts. LandingAI states that Verity uses roughly 40% of the credits charged by Pro, while automated routing between the models is planned for autumn 2026. Source: Marktechpost
The new system also changes how document processing is measured. Instead of charging a flat amount for every page, DPT-3 combines a page component with an output-character component. LandingAI says priority processing is designed for situations where a person or software agent is waiting, while standard processing runs asynchronously for workflows that can tolerate a longer processing window. Source: Marktechpost
“Atomic grounding” means an extracted answer can be connected to a specific visual line or word on a specific page, rather than simply appearing as unsupported text.
Gen2’s response contains markdown, metadata and a structure tree. Each block includes a semantic identifier and grounding information, including its page, position in the output and normalised coordinates. DPT-3 Pro provides grounding at visual-line level, while DPT-3 Verity provides word-level grounding and confidence information. Source: Marktechpost
Why This Matters for Malaysian SMEs
Many Malaysian businesses work with documents that are not neatly formatted. A wholesaler may receive supplier invoices in PDF, image and spreadsheet formats. A renovation company may photograph signed work orders at a job site. A logistics operator may process delivery orders with stamps, handwritten quantities and partial signatures. A clinic, tuition centre or professional services firm may need to extract details from registration forms and supporting documents.
Traditional OCR can turn an image into text, but plain text often loses the relationship between a value and its label. A total amount may be separated from the invoice number. A table may become a confusing sequence of lines. A signature may be mistaken for ordinary marks. Gen2’s block-based structure is intended to preserve these relationships, making it easier for your automation system to distinguish an invoice table from a note, a stamp from a printed label, or a figure from surrounding text.
This can support common SME workflows such as:
- Accounts payable: Extract supplier name, invoice number, date, tax details and line items before sending them for approval.
- Inventory updates: Read quantities and product codes from delivery notes, then compare them with purchase orders.
- Sales administration: Capture customer information from forms without requiring staff to key in every field.
- Contract review: Locate renewal dates, obligations and signatures while keeping a reference to the original page.
- Compliance checks: Flag unclear or low-confidence words for human review rather than silently accepting them.
The confidence score from Verity is particularly useful for a small team. You do not need to check every character equally. A workflow could automatically accept clear printed invoice numbers, while sending a blurred handwritten bank account number or unclear quantity to a staff member. The result is not the removal of human judgement; it is better allocation of your team’s attention.
Key Capabilities to Watch
| Capability | Practical SME use |
|---|---|
| Document and page structure | Keep tables, figures, signatures and text in their correct context. |
| Word-level confidence | Identify uncertain transcription for manual checking. |
| Atomic grounding | Show the original page and location supporting an extracted field. |
| Table-cell coordinates | Support more accurate review of quantities, rates and totals. |
| Multiple deployment options | Consider cloud, private environments or on-premises processing for sensitive records. |
| Asynchronous processing | Run document pipelines in the background when immediate results are unnecessary. |
These capabilities are described in LandingAI’s Gen2 release coverage. Source: Marktechpost
What You Should Do Before Adopting It
Start with one document type rather than attempting to automate every administrative process at once. Choose a repetitive workflow with a clear success measure, such as extracting fields from supplier invoices or matching delivery notes to purchase orders. Collect examples covering different suppliers, layouts, languages, scan quality and handwritten annotations.
Next, define which fields require human approval. For example, your system might automatically capture a supplier address but require review for tax numbers, bank details, unusual quantities and contract commitments. This creates a controlled workflow where automation handles predictable work and your staff remains responsible for important decisions.
You should also test how the character-based billing model behaves with your own documents. LandingAI states that DPT-3 processing combines page and output-character components, so dense documents may behave differently from short forms. The company also publishes vendor estimates for reductions and per-page processing, but you should treat those figures as claims to validate with your own document mix. Source: Marktechpost
The Bigger Picture
The important development is not simply that software can read more document types. It is that document extraction is moving towards auditable automation. When an extracted field includes its page and location, your team can review the source instead of trusting an isolated answer. That is valuable for internal controls, customer disputes, supplier reconciliation and management reporting.
For Malaysian SMEs, this could make document automation more practical because many businesses operate with a mixture of digital PDFs, mobile photos, scanned copies, stamps and handwritten records. The best implementation will not be a system that tries to replace every administrative role. It will be a workflow that captures routine information, highlights uncertainty and gives your people a fast way to verify the result.
Keep in mind that Gen2 is a new interface and Gen1 client code will not run against Gen2 endpoints, according to the release coverage. Any implementation should therefore include migration planning, data protection checks, user permissions, retention rules and testing against your actual documents. Source: Marktechpost
Your next step is simple: select one document-heavy process, measure how long it currently takes, identify the fields that cause the most errors and test whether grounded extraction can make review faster and more reliable. For a small business, that focused approach is more useful than adopting technology merely because it is new.
Ready to Streamline Your Operations?
Technology moves fast. Your operations should keep up. AutoRunBiz builds AI systems that run your daily workflows — from WhatsApp order capture to accounting. Book a free 15-min ops audit →