500,000 certificate and career documents a year, entered by hand — could AI take over? An on-premise AI-OCR validation

Summary

A public institution in Korea asked us whether the manual processing of roughly 500,000 certificate and career documents a year, handled by six staff members, could be automated. We built a workflow in which AI extracts the key fields from each document, assigns a confidence score to every extracted item so reviewers check the least certain ones first, and auto-approves entries once confirmed. Because the documents contain personal data, the entire build and test ran on-premise, inside the institution's internal network.

The problem — 500,000 documents, six people, all typed by hand

• Roughly 500,000 supporting documents — licenses, certificates, career records — arrive every year
• Six staff members read each document and type the key fields (qualification name, issuing body, acquisition date, employment period) into the system by hand
• That is more than 80,000 documents per person per year: delays accumulate, and repeated manual entry brings typos and omissions with it

The validation question was simple: can this run without manual typing?

Why plain OCR wasn't enough

Standard OCR turns pixels into text — but it has no idea which text is the qualification name and which is the acquisition date. Certificates come in different layouts from every issuing body, with fields buried inside tables, so text extraction alone still leaves a human to classify and re-type everything.

Why ROI (fixed-coordinate) templates failed too

The common approach in document automation is ROI (Region of Interest) templating: define "the qualification name lives at these coordinates on this form" and read that region. It works well when layouts are fixed.

These documents were not like that.

• Even within the same form, field positions shifted from page to page. A longer entry higher up pushes everything below it further down
• Line-breaking differed from document to document. A career entry that fits one line in one document wraps onto two lines in another

Read by coordinates, some documents parse correctly and others return the wrong values entirely. Even maintaining a template per form couldn't absorb this variance — these were effectively unstructured documents, beyond what ROI-based or plain OCR could automate.

Reading by meaning, not position

SUVsoft AI-OCR reads the key-value relationships inside a document instead of coordinates: "qualification name → Engineer X", "acquisition date → 20XX.XX.XX". Because fields are identified by meaning, the extraction result stays the same whether an item shifts down the page or wraps onto a second line — and it comes out in a form the business system can ingest directly.

The workflow we validated — extract, confirm, auto-approve

1. Automatic extraction — AI pulls the fields the screening process needs: qualification name, issuing body, acquisition date, employment period
2. Confidence-based review — the reviewer's screen shows the original document image and the extracted values side by side. Staff confirm rather than type
3. Auto-approval — once confirmed, the entry is approved automatically and flows into the downstream process

The staff member's role changes from "read and type" to "review and approve". Since manual entry disappears from the flow, entry typos and omissions are eliminated structurally.

If humans must check everything, nothing is saved — so confidence sets the order

Automatic extraction is pointless if staff still have to eyeball all 500,000 documents; the work has merely been renamed from "entry" to "review". This is where document automation projects typically fall short of expectations after deployment.

So the engine cross-checks each extracted item internally and computes a confidence score. The review screen surfaces low-confidence items first, letting staff spend their time where checking is actually needed.

Reviewers can still open and inspect every item — the system doesn't hide results behind the AI's judgment; it proposes an order to look at them. Final decisions in a screening process belong to people. Confirmed entries proceed to auto-approval.

Privacy — on-premise build with automatic PII masking

The documents carry names, dates of birth, and Korean resident registration numbers. Sending them to an external cloud API was never on the table.

• On-premise validation — the entire build and test ran inside the institution's internal network. Neither documents nor extracted data ever left it
• Automatic PII detection and masking — SUVsoft's PII detection technology identifies resident registration numbers and similar identifiers in the documents and masks them, so personal data is not exposed unnecessarily during processing

What the validation confirmed

• Key-value extraction works on real document types whose layouts differ by issuer — and shift even within the same form — without relying on coordinates (ROI)
• Per-item confidence scoring works, and ordering the review queue by low confidence works as designed
• The full flow — extraction → side-by-side confirmation → auto-approval — is implementable entirely on-premise
• Automatic detection and masking of resident registration numbers and similar PII works on the actual documents

This case describes a technical validation (PoC). If the project proceeds to production, we will update this page with operational results.

Frequently Asked Questions

Q. How is this different from a regular OCR product? Regular OCR converts images to text. It cannot tell which text is the qualification name and which is the acquisition date, so a human ends up classifying everything anyway. AI-OCR extracts key-value pairs — data your system can ingest directly.

Q. Layouts keep changing — do we have to reconfigure every time? With coordinate (ROI) templates, yes. Key-value extraction reads the document's semantic structure rather than positions, so it is considerably less sensitive to layout changes.

Q. Don't humans have to check everything anyway? Review is still part of the process — but each item carries a confidence score, and the queue surfaces the least certain items first. That is a very different time budget from scanning everything with equal attention.

Q. The documents contain personal data. Does anything leave our network? Deployed on-premise, neither documents nor extracted results leave your internal network. Automatic detection and masking of resident registration numbers and similar identifiers can be applied as part of the pipeline.

Environment

Related product: AI-OCR — Intelligent Document Processing

Have a similar workload? If people in your organization read documents and type them into a system at scale, the same approach applies — whatever the document type. Tell us about your situation and we will start with a feasibility review.