11. AI Classification
Teach Ada your own document types and use them in rules.
Policy → AI Classification
AI classification recognises what kind of document a file is from its meaning, not from the words it contains: board minutes, an employment contract, a CV, a customer offer. You teach Ada your own document types with example documents; the server trains a small decision layer; agents score documents locally and the result is used as a condition inside ordinary Classification Rules.
This chapter covers the AI page and its sub-tabs, training and how to read its results, AI conditions in rules, delivery of the model to endpoints, cost, privacy, rollout and troubleshooting.
In this topic
- 11.1 Semantic
classification versus pattern rules
Key points:
- 11.2 Components and
prerequisites
Before AI does anything on an endpoint, all of the following must be true:
- 11.3 The AI Classification page
The page has a collapsible help box ( How does AI classification work? ) and three sub-tabs:
- 11.4 Overview
Policy → AI Classification → Overview shows two cards.
- 11.5 Your document types
Policy → AI Classification → Your document types is a single card that holds the whole teaching cycle:
- 11.6 Training
Training usually takes seconds; it runs in a background thread on the server and does not slow the console or agent traffic. It is stopped after 15 minutes.
- 11.7 Using AI in
Classification Rules
The AI module is off by default and is the only module whose default is off.
- 11.8 Test a document
Policy → AI Classification → Test a document scores one file exactly as the agents read it.
- 11.9 How the model reaches
the endpoints
The server adds two things to the signed policy it serves to agents (they are not in the editable policy and cannot be changed in the console):
- 11.10 Where AI runs and what it
costs
A per-user AI worker (its own launch agent, standard scheduling class, lowered priority) scores for that user’s file agent; if it is unavailable, the file agent scores in-process (slower, same result). The root…
- 11.11 What users see and
what is audited
Label window (Windows Office and right-click Classify). When an AI rule recommends a label, the window opens with that label pre-selected and a line such as:
- 11.12 Privacy
Everything runs on-premises. No document, text or score is sent to any cloud or AI service; ada-ml has no route to the internet. Endpoints send no text. Agents report only the document type, probability and rule —…
- 11.13 Recommended rollout
Tip Combine AI with patterns for high-confidence enforcement, for example Match: All of “AI: Payroll ≥ 80 %” and “pattern: national ID number”. Both must hold, which sharply reduces false positives.
- 11.14 Troubleshooting
The Windows service also writes a ten-minute summary ( AI scoring in the last 10 min: … ) to service.log ; it never contains document text.
- 11.15 Reference: limits and
defaults
Item Value AI module Off (Default and new groups) Document types Max 50; name 2–60 characters, unique Examples Min 5 per type to train; 20+ recommended; max 500 per type; max 3,000 total File size (examples, test…