How Classifyre Works

This section explains Classifyre the way an operator or investigator would describe it — not a developer. If you want the full technical detail on any piece, each section links to the deeper docs.

At its core, Classifyre does one thing: it turns the data scattered across the systems you already run into a small number of leads worth a human’s time, and gives you a structured place to work them.

Prefer to learn screen by screen? Take A Tour of the App — it maps every item in the app’s navigation to what it’s for.

Everything below happens inside a workspace — an isolated investigation with its own sources, findings, cases, and AI memory. One Classifyre instance can hold many of them (a client, a region, a business unit), and nothing is shared between them.

The journey, in plain English

1. Connect a source. Point Classifyre at a system you already run — SharePoint, Confluence, Jira, a file share, a local folder. Nothing is moved out of your control; Classifyre reads it in place.

2. Scan it. A scan crawls the source and registers every document or file as an asset, with metadata like owner, location, and last-changed date.

3. Detect. Detectors — pattern rules, ML models, or AI-based custom detectors — read each asset and raise a finding whenever something’s worth a look, each carrying a severity (how bad if real) and a confidence (how sure the detector is).

4. Rank. Severity alone doesn’t say what to look at first. A ranking pass scores each finding’s importance — how much it looks like a genuine lead, independent of severity — using signals like recurrence across documents, context quality, and whether it looks like test data. See Ranking & the Semantic Layer.

5. Watch and connect. Inquiries are saved questions that keep watching for matching findings; fingerprints link assets sharing concrete values like an email or ID — revealing when the same record shows up in several systems. See Connections & Fingerprints.

6. Investigate. A case is where the real work happens. Promising findings arrive as leads — a triage queue you accept into evidence or dismiss. You record dated events to build a chronology, and write a conclusion when resolved. See Leads, Evidence & Events.

7. Keep a shared vocabulary. A glossary of people, organisations, and terms keeps everyone — human and AI — talking about the same entities. See Glossary & Shared Vocabulary.

8. Let AI help. Autopilot agents can do this legwork for you — triaging findings, proposing leads and events, tuning detectors — always logging why, and always subject to the supervision level you choose. See Autopilot & AI Assistance.

Why it’s built this way

Two ideas run through the whole product:

• Severity is not importance. Severity says how bad a match would be if real. Ranking says how much this finding, in context, looks like a genuine lead. A “critical” finding repeated as boilerplate everywhere isn’t where you start.
• Similarity is not proof. Two things looking alike — semantically or by shared values — is a reason to look, not to conclude. Every automated suggestion, from a ranked finding to an Autopilot-proposed lead, is a candidate for a human to accept or dismiss.

Where to go next

Technical reference: Scans, Sources, Detectors, Investigations.