Introduction. What Cogneris is, in plain terms.

What Cogneris does

Documents in, structured decisions out — with the work shown.

• Document AI, not OCR — agents reason over the page, ground claims to pixels, and admit when they don't know.
• Structured output — typed JSON against a template or an inline schema; every field carries confidence and a bounding box.
• Decisioning, not just parsing — pair extraction with validation, fraud scoring and custom agents to actually route the work.
• Audit trail by default — every run is replayable: prompts, tool calls, retrieved passages and the final answer.

Mental model

Three layers, top to bottom.

The five agents

A typical flow

Vendor invoice processing, end to end.

1. Classify — the inbox blob is an invoice (vs PO, vs packing slip). 1 API call.
2. Extract — pull vendor, PO, line items, tax, total against the invoice template.
3. Validate — totals must reconcile to ±0.01; tax ID must match the regex; PO must link to a known purchase order.
4. Score for fraud — tampering, font inconsistency, metadata anomalies; route 70+ to a second reviewer.
5. Hand off — push the row to your ERP, or kick a webhook to your workflow tool.

Each of those is a single endpoint call — or a single Custom-agent run that chains all five.

Multi-tenancy & audit

• Tenant-scoped data — every row carries tenant_id; cross-tenant reads are physically impossible by design, not by convention.
• Audit trail — every run stores prompts, tool calls, retrieved passages, retries and the final decision. Surfaced as an audit_url on every response.
• Soft deletes — deletions flip an IsDeleted flag and stamp DeletedWhen; nothing is purged silently.
• Region pinning — choose where your tenant lives at sign-up; data stays there unless you explicitly move it.

When to reach for what

• "I just need fields off a PDF" — Extraction with a shipped template.
• "My inbox has mixed document types" — Classification first, then route to extraction.
• "I have one big document and human questions" — Ask the PDF.
• "I need to keep bad ones out" — Fraud detection, paired with Validation rules for deterministic checks.
• "I want all of the above as one workflow" — Custom agents.