Glossary
Your Guide to the Agentic Era.
Agentic AI, ontologies, RAG, LLMs — the technologies and ideas driving AI in the enterprise, explained plainly.
- Terms
- 54
- Categories
- 5
Core concepts
10AI Worker
AI WORKER
An AI that reads and structures documents and data, makes judgments and takes over work people did, up to entering the result into the business system.
DetailsZero-Touch
ZERO-TOUCH
A setup where AI finishes normal cases and people review only exceptions, cutting cases people touch. The 3 Zero AI Worker strategy calls it Zero Review.
DetailsUnstructured Data
UNSTRUCTURED DATA
Documents, images, faxes and scans that are not in table form and cannot go into a system as is. Most of the data an organization holds is of this kind.
DetailsDocument AI
Document understanding technology
Document AI is the category of AI technologies that recognize and understand documents. It covers reading; work done after reading is called AI Worker.
DetailsIDP
INTELLIGENT DOCUMENT PROCESSING
IDP links document intake, recognition, extraction and validation into one flow. It has to be designed around the business process to cut real workload.
Details3 Zero AI Worker
3 ZERO AI WORKER
The 3 Zero AI Worker strategy builds document automation on Zero Training, Zero Hallucination and Zero Review, judged by the share handled without people.
DetailsDocument Automation
DOCUMENT AUTOMATION
Document automation is AI taking over opening documents, checking them and re-keying values into systems. Just displaying results adds a checking step.
DetailsDEEP Agent
DEEP AGENT
DEEP Agent, Korea Deep Learning's AI agent, understands unstructured data, judges by business rules and context, and carries out review and registration.
DetailsAX
AI TRANSFORMATION
AX (AI Transformation) is an organizational shift in which AI performs part of people's work, changing workflows, systems and operations together.
DetailsDocument AX
DOCUMENT AI TRANSFORMATION
Document AX applies AX to receiving, reading, judging and registering documents, so AI runs the whole process that people used to stitch together.
Details
Work and processes
8RPA
ROBOTIC PROCESS AUTOMATION
RPA software repeats fixed screen operations and data entry. It works best where rules are fixed, so AI commonly reads documents and RPA enters the data.
DetailsEDMS
Electronic Document Management System
An EDMS stores and searches document originals and processing history. Extracted data must land in it before RPA and core systems can continue the work.
DetailsSTP
NO-INTERVENTION RATE
STP is the share of received cases finished end to end with no human touch. It shows real labor savings more accurately than recognition accuracy.
DetailsHuman-in-the-Loop
Human Review Intervention
A design in which people check only the AI results with low confidence. It is a practical way to manage operational risk between full review and no review.
DetailsException Handling
EXCEPTION HANDLING
Exception handling routes cases that fixed rules cannot judge to a separate path. Automation succeeds or fails on how narrowly exceptions are defined.
DetailsFTE Reduction
FULL-TIME EQUIVALENT
Converts the staff needed for the same workload into full-time equivalents. It is the figure most often requested to explain document automation's return.
DetailsTask Completion Rate
TASK COMPLETION RATE
The share of cases finished within the workflow, through judgment, system registration and result recording. Check it if staff still re-enter results.
DetailsAgentic Document Processing
AGENTIC DOCUMENT PROCESSING
Document processing where an AI agent decides what to check and which tools to use, within set goals and permissions, and carries the work to the end.
Details
Technology and architecture
13OCR
OPTICAL CHARACTER RECOGNITION
OCR converts text in images into machine-readable text. Business documents are harder than printed type: handwriting, tables and stamps share one page.
DetailsVLM
VISION LANGUAGE MODEL
A VLM understands images and text together, reading layout and context along with characters. Gaps show on merged cells, checkboxes and poor-quality faxes.
DetailsDocument Parsing
Document structuring
Document parsing turns headings, tables, lists and figures, with their hierarchy, into machine-usable data. RAG and search quality is decided here.
DetailsKIE
KEY INFORMATION EXTRACTION
KIE extracts only the items a business needs from a document as key-value pairs. OCR answers what is written; KIE answers what it means for the business.
DetailsLayout Analysis
LAYOUT ANALYSIS
Layout analysis separates a page's titles, body text, tables and headers and sets their reading order. Accuracy gaps show in multi-column layouts.
DetailsDocument Classification
DOCUMENT CLASSIFICATION
Document classification identifies a file's form type and splits it into documents. When documents arrive merged in one PDF, this step is the bottleneck.
DetailsTable Recognition
TABLE RECOGNITION
Table recognition moves a table's rows, columns and merged cells into data with structure intact. Without structure, items lose their link to amounts.
DetailsCross-Document Verification
CROSS VALIDATION
Compares the same item across the documents in one case to judge whether they match. If names or account numbers conflict, the work cannot proceed.
DetailsAgent Workflow
AGENT WORKFLOW
An automation design that links classification, extraction, validation, judgment, approval and execution into one flow, so each case reaches the end.
DetailsAI Agent
AGENT
AI that takes a goal, picks its own tools and chains several steps. In document work it judges conditions on extracted data and registers the result.
DetailsExtraction Schema
EXTRACTION SCHEMA
An extraction schema defines which items to pull from a document, under what names, formats and conditions, so business systems can use the results as is.
DetailsAgentic AI
AGENTIC AI
AI that, given a goal, plans its own steps, picks and runs tools, and adjusts from the results. Consider it where people still chain lookups by hand.
DetailsLLM OCR
LARGE LANGUAGE MODEL OCR
LLM OCR reads documents with a large language model that uses context to read characters and fill required fields. It can invent values, so cross-check.
Details
Data and quality
10Exact Match
Exact match rate
The share of extracted values identical to the correct answer, character for character. It is the strictest standard, so set operating targets separately.
DetailsFuzzy Similarity
Similarity-Based Accuracy
Measures how well two strings match, allowing for spacing and symbol differences. It reflects what people read as the same value, near perceived accuracy.
DetailsTEDS
TABLE STRUCTURE SIMILARITY
TEDS measures how closely an extracted table matches the original's structure. Even with high text accuracy, a low score means misaligned table data.
DetailsPersonal Information De-identification
DE-IDENTIFICATION
De-identification masks or replaces items that can identify a person. It is effectively a prerequisite for using public data in training or search.
DetailsFine-Tuning
FINE-TUNING
Fine-tuning trains a pre-trained model further on specific documents and business rules. On fixed-format documents, the gap with general models is large.
DetailsHallucination
HALLUCINATION
A model producing plausible values without evidence. An invented amount or date is a business incident, so return each value's evidence location.
DetailsGround Truth Data
GROUND TRUTH DATA
Human-confirmed correct values for each document item or question, used to score AI results and compare versions. Build it before judging changes by eye.
DetailsData Drift
DATA DRIFT
Data drift is when data entering a live system no longer matches what the model was built and evaluated on. It slowly lowers a working AI's performance.
DetailsConfidence Score
CONFIDENCE SCORE
A per-case number showing how likely an AI result is to be correct. It separates cases to process automatically from cases a person should check.
DetailsAI-Ready Data
AI-READY DATA
Data organized with structure, sources, quality and usage terms so AI can use it directly. If AI answers blur tables or lack sources, this step is missing.
Details
Adoption and operations
13AI Ops
AI OPERATIONS
The operating system for AI models and agents after deployment, managing quality, cost, latency, versions and change history, prompts and tools included.
DetailsDEEP Ops
DEEP OPS
DEEP Ops is Korea Deep Learning's document AI operations platform. Teams change extraction criteria as work changes, then validate and version them.
DetailsMLOps
MACHINE LEARNING OPERATIONS
MLOps manages a machine learning model across development, deployment, monitoring and updates, focused on keeping performance stable in real use.
DetailsOn-Premise
ON-PREMISE
On-premise means running models and systems on the customer's own servers, not an external cloud. In finance and the public sector it is a requirement.
DetailsNetwork Separation
NETWORK SEPARATION
A Korean security regulation that splits the work network from the internet. External API calls are blocked, which limits the solutions you can adopt.
DetailsOCR API
OCR API
An OCR API returns recognition results for document images you send. It suits early validation; with security requirements, check for an on-premise option.
DetailsRAG Preprocessing
RAG PREPROCESSING
RAG preprocessing organizes documents for retrieval-augmented generation, keeping structure and meaning. Poor answers often start here, not in the model.
DetailsChunking
CHUNKING
Chunking splits a long document into pieces sized for search and generation. Cutting by character count alone breaks tables and clauses in the middle.
DetailsVLMOps
VISION LANGUAGE MODEL OPERATIONS
VLMOps keeps models, prompts, extraction schemas and quality standards under version control and evaluation. It lowers upkeep where forms change often.
DetailsBMT
BENCHMARK TEST
BMT compares candidate solutions under identical conditions on your real work documents. Public benchmark scores differ, so base adoption on BMT results.
DetailsHigh-Impact AI
HIGH-IMPACT AI
Under the AI Basic Act, AI that may seriously affect people's safety or fundamental rights. Providers owe risk management, explanation and human oversight.
DetailsAgentOps
AGENT OPERATIONS
Operations for AI agents that act in business systems: recording each action and managing permissions, approvals, per-task evaluation, stop and recovery.
DetailsLLMOps
LARGE LANGUAGE MODEL OPERATIONS
LLMOps manages prompts, model versions, search data, answer quality and cost together to keep large language model results consistent after deployment.
Details




