06

Cemetery Intelligence: Exception Detection & Review Engine

Exception Detection & Review Engine

A rules-first review workflow that normalises records, applies deterministic checks, investigates uncertain findings with AI, and requires human approval before a finding becomes client-ready. Nothing reaches a client until a person approves it.

Class
AI Operations
Built with
n8n · OpenAI · Google Sheets · Deterministic rules · Confidence scoring
Walkthrough
None recorded

The work

Records need to be normalised, checked against several business-rule categories, investigated where needed, and reviewed before any finding can be shared.

The friction

The repetitive work is preparing records, checking several rule categories, merging findings, determining confidence, deciding which cases deserve deeper investigation, and organising findings for human review.

n8n workflow canvas for Cemetery Intelligence, showing record parsing, normalisation and indexing, three parallel deterministic rule paths for burial and monument, ownership and customer, and documentation and work, a merge and confidence scoring step, an AI investigation chain using an OpenAI chat model, and a human review gate that writes to a Google Sheets review queue. A separate branch shows the human decision trigger and the approved path that finalises a client-safe finding.
The full n8n canvas. The three deterministic rule paths run in parallel on the left, AI investigation sits in the middle, and the run terminates at the review queue. The approved path on the separate branch only executes after a person decides.1863 × 811 · 180 KB

The system

Deterministic rules run first and in parallel. AI investigates only what the rules leave uncertain, and the run stops at a review queue.

  1. Run Synthetic Cemetery Demo
  2. Workflow Configuration
  3. Synthetic CSV Test Data
  4. Parse, Normalize and Index Records

  5. Deterministic rule paths, run in parallel

    • Rules: Burial and Monument
    • Rules: Ownership and Customer
    • Rules: Documentation and Work
  6. Merge All Findings
  7. Deterministic Confidence Scoring
  8. Prepare AI Investigation
  9. AI Investigation Chain, using an OpenAI chat model
  10. Parse AI Investigation JSON
  11. Assemble AI Investigation
  12. Human Review Gate, Prepare Queue
  13. Append to Google Sheets Review Queue

Human decision flow, a separate run

  1. Human Review Decision Trigger
  2. Is Human Decision Approved?

    Only the approved path continues

    • Approved: Finalize Approved Client-Safe Finding
    • Approved: Append Approved Finding to Client-Ready Google Sheet

What it handles

  • Synthetic and test record ingestion
  • Normalisation
  • Indexing
  • Deterministic rule evaluation
  • Multi-rule finding aggregation
  • Confidence scoring
  • AI investigation preparation
  • AI-assisted investigation
  • Structured JSON parsing
  • Review-queue preparation
  • Google Sheets review queue
  • Post-approval finalisation

What stays human

Human review is mandatory. A finding sits in the review queue until someone approves it, and only approved findings continue to the client-ready output.

  • Approving or rejecting each finding
  • Judging whether the evidence is sufficient
  • Deciding what is safe to put in front of a client
  • Everything downstream of the review gate

The principle

Automation prepares the decision. A person still makes it.

Exceptions

This system is built around exceptions. Surfacing them is the point of the workflow, not a side effect.

  • Incomplete records
  • Conflicting records
  • Inconsistent findings
  • Low confidence
  • Disagreement between deterministic rules
  • Uncertain AI findings
  • Cases that should not proceed automatically

How it improves the workflow

The workflow separates repetitive record analysis from the judgment required to approve a client-facing finding. Rules and AI prepare the evidence, and the review gate stays in place.