Understanding AI classification

How the Classifier Agent assigns categories + tax tags, how your corrections train it, and how to reach 95%+ accuracy in 3 weeks.

In this guide

1. What gets classified

Every transaction flowing into HaraPro — whether via Plaid or PDF statement — gets passed to the Classifier Agent. The agent looks at:

It then assigns one category (e.g., "Meals & Entertainment") and up to 3 tax tags (e.g., "Deductible", "Business Meal 50%", "Travel-related"). Each assignment gets a confidence score 0–100.

2. The 49 categories

Categories are the "what is this?" layer. We use 49 that map cleanly to IRS Schedule C, Schedule E, and common S-Corp and partnership accounts:

Revenue / Sales
Cost of Goods Sold
Contractor Fees
Professional Fees
Legal Fees
Accounting
Advertising
Marketing
Software & SaaS
Office Supplies
Rent — Office
Rent — Equipment
Utilities
Internet / Phone
Insurance — Business
Insurance — Health
Vehicle Expense
Vehicle Depreciation
Meals & Entertainment
Travel
Dues & Subscriptions
Bank Fees
Interest Expense
Loan Principal
Credit Card Payment
Owner Draw
Owner Contribution
Payroll — Wages
Payroll Taxes
Benefits
Retirement Contributions
Education
Charitable Contributions
Taxes — Federal
Taxes — State
Taxes — Property
Repairs & Maintenance
Depreciation
Amortization
Investment Income
Investment Expense
Rental Income
Rental Expense
Distribution
Transfer
Refund / Reversal
Personal — Living
Personal — Medical
Uncategorized

You can rename categories per entity (e.g., "Software & SaaS" → "Software — Dev Tools") and create custom sub-categories.

3. The 17 tax tags

Tax tags are the "how does this affect taxes?" layer. They stack on top of categories and drive the tax forecast, deduction simulator, and CPA export. The 17:

Deductible
Partially Deductible
Non-Deductible
Capital Expense
§179 Eligible
Bonus Depreciation
§280F Limited
Business Meal 50%
Home Office
Vehicle — Mileage
Vehicle — Actual
Self-Employment Tax
Retirement Contribution
HSA Contribution
Charitable
Personal (no deduction)
Under Review
Tip: Tax tags are where the real leverage lives. A $65,000 SUV is a "Vehicle Expense" category — but the tax tag "§179 Eligible" (if over 6,000 lbs GVWR) could mean a $65K deduction in year 1 instead of $13K. The Deduction Simulator in Pro+ lets you model this before you buy.

4. The 4-layer memory system

The Classifier doesn't learn in isolation — it uses four layers of memory, stacked in priority order (top wins):

1

Session context

The immediate context of the transaction under review — adjacent transactions, current entity, recent batch uploads. This resolves ambiguity for things like "Amazon" appearing 40 times in a week (all business supplies vs mixed personal).

2

User corrections (your tenant)

Every time you override a classification, that correction is stored. The next time the same merchant (or similar description) appears in the same entity, your prior correction wins. Corrections propagate within 24 hours.

3

Tenant preferences

Patterns specific to your HaraPro account: "This tenant tags all Shell purchases as Vehicle — Actual, not Personal" or "This tenant categorizes Uber consistently as Travel for entities A, B and Personal for entity C." Tenant preferences apply across all entities in your account.

4

Global patterns (anonymized)

Industry-wide patterns from all HaraPro users, fully anonymized. "Intuit" is overwhelmingly "Software & SaaS" in the SMB segment. "Delta Airlines" is overwhelmingly "Travel". This is the fallback when your own tenant has no prior signal.

5. How to correct classifications

Open the Transactions tab. Low-confidence classifications land in Needs review automatically. To correct:

  1. Click any transaction row
  2. Click the Category dropdown to change it
  3. Click Tax tags to add/remove tags
  4. Optionally: Apply to all similar — to retroactively fix every matching past transaction

The "Apply to all similar" option uses merchant name + description fuzzy match. Useful when you realize "Aws" has been miscategorized as "Uncategorized" for months — one click fixes all 47.

6. Reaching 95%+ accuracy

Typical accuracy trajectory for a new tenant:

To accelerate: when you do your first review, go through 200–300 transactions in one sitting. The AI learns your patterns much faster from a concentrated burst than from scattered fixes over months.

Privacy note: Your corrections only train your tenant. They never leak to other users. Global pattern learning uses anonymized, aggregated signals only — never individual transaction text. AI processing runs under enterprise no-training agreements with OpenAI and Anthropic. Read the privacy policy →