How AI Categorization Actually Works in EXPOZOR

Rules first, AI second

Most "AI-powered" finance apps apply machine learning to every transaction. That sounds impressive until the AI confidently labels your gym membership as "Entertainment" for the third month in a row.

EXPOZOR takes a different approach: your rules always run first.

When a new transaction arrives — whether from a bank sync, a receipt scan, or manual entry — the system processes it in this order:

1. Exact merchant rules — You said "Planet Fitness = Health & Fitness"? Done. No AI involved.
2. Pattern rules — Any transaction containing "AWS" goes to "Business / Infrastructure"? Applied instantly.
3. AI categorization — Only if no rule matches does the AI step in.

This means the AI is a fallback, not a gatekeeper. And it knows its place.

Confidence scoring

When the AI does categorize a transaction, it assigns a confidence score between 0 and 100:

| Confidence | Behavior | |-----------|----------| | 80–100% | Auto-applied. You see a small "AI" badge on the category. | | 60–79% | Auto-applied with a "Review" nudge in your activity feed. | | Below 60% | Sent to your review queue. Never auto-applied. |

The threshold is configurable. If you want the AI to be more cautious, you can raise the auto-apply threshold to 90%. If you trust it more, lower it to 50%.

Explainability

Every AI-categorized transaction comes with a short explanation:

"Categorized as 'Dining Out' because the merchant 'Nando's' matches 847 similar transactions in our training data, and the amount ($23.50) is within the typical range for restaurants."

You can tap the explanation to see the full reasoning chain. If the AI got it wrong, you correct it once, and it creates a new rule automatically — so it never makes the same mistake again.

The training data question

A fair question: where does the AI's training data come from?

• Aggregate patterns — anonymized, de-identified transaction patterns from public financial datasets.
• Your corrections — when you fix a categorization, it improves your model. Your data never trains anyone else's model.
• No selling, no sharing — your transaction data is encrypted with per-user envelope keys. We literally cannot read it.

Why this matters

Categorization accuracy isn't a vanity metric. It's the foundation of everything else: budgets, spending reports, split calculations, and financial insights.

If your categories are wrong, your budget is a fiction. EXPOZOR's rules-first approach means your categories are right from day one — and they only get better.