AI should assist, not decide

Inzolo uses AI — carefully. Here’s the rule we won’t break: rules and your own history come first, AI is a fallback, and nothing touches your money without your approval.

It would be easy to slap “AI-powered” on a budgeting app and call it the future. Point a model at your transactions, let it categorize everything, and ship. Plenty of apps do roughly that.

We didn’t, for one stubborn reason: your money is not a good place for a confident guess. A model that’s right 95% of the time is genuinely useful — and also quietly wrong about one transaction in twenty, every single month, in ways you might not notice until your budget no longer matches reality. So we built the categorization system in a specific order, with AI deliberately near the bottom.

The pipeline, in order

When a transaction lands, Inzolo works through a sequence and stops at the first confident answer:

  1. Normalize the messy payee string (SQ *BLUE BOTTLE #441 → “Blue Bottle Coffee”).
  2. Your explicit rules — if you’ve said “anything from this payee is Groceries,” that’s the answer. Deterministic, instant, yours.
  3. Your history — how you’ve categorized this payee before. The system you’ve already trained, just by using it.
  4. Heuristics for the obvious cases — transfers, fees, recurring subscriptions.
  5. AI — only when the steps above are uncertain, and only to suggest, always with a confidence score and a short explanation.

Notice that AI is step five, not step one. By the time a question reaches the model, the cheap, predictable, fully-explainable methods have already handled the easy 80%. The model gets the genuinely ambiguous remainder — which is exactly what it’s good for.

Deterministic rules and your approved history are the core system. AI is the careful fallback for the uncertain cases — never a hidden hand on your money.

Nothing moves without you

This is the line that matters most. Every imported transaction — whether it was categorized by a rock-solid rule or a low-confidence AI guess — lands in the approval queue as unapproved. It does not touch an envelope balance. It does not change your “safe to spend.” It waits.

You see the queue sorted by confidence: the sure things at the top for a quick glance-and-approve, the “needs you” items flagged honestly at the bottom. You approve, recategorize, or split. Only then does it become part of your budget. There are no silent corrections happening while you sleep.

Learning that stays legible

When you approve a transaction, Inzolo can quietly offer to turn that decision into a rule — so next month it’s handled automatically, by a rule you can read and edit, not by a black box that drifts over time. The system gets smarter in a direction you can always inspect and undo.

We keep an audit trail on suggestions too: which method produced it, the confidence, the model and prompt version, and who approved it. If something ever looks off, you can see exactly how it got there.

Where AI genuinely shines

None of this is anti-AI. It’s about pointing the tool at the right jobs — the ones where a thoughtful suggestion beats a rigid rule:

  • A plain-language weekly coaching summary of what changed and what to watch.
  • Explaining a confusing number in a sentence instead of a formula.
  • Mapping the columns of an unfamiliar CSV during import setup.
  • Spotting a likely recurring bill and offering to set a reminder.

The principle, in three words: AI assists, rules decide, you approve. It’s the only arrangement we’d trust with our own household’s money — so it’s the one we built.

← Back to the blog Create account