AI Can Not Replace Judgment: Edge Cases That Still Need Human Categorization

AI categorization gets most transactions right. But “most” is not “all,” and the edge cases reveal why human judgment remains essential. Here is my collection of transactions that consistently confuse automated systems.

The Classic Edge Cases

The Ambiguous Merchant

AMAZON - Is it:
- Books (entertainment)?
- Office supplies (business)?
- Groceries (food)?
- Electronics (equipment)?
- Returns (negative expense)?

AI cannot know without seeing the receipt.

The Split Transaction

TARGET $127.34 - Actually:
- $45.00 groceries
- $32.00 household supplies
- $25.00 clothing
- $25.34 pharmacy

One transaction, four categories. Automation cannot split without the receipt.

The Reimbursable Expense

DELTA AIRLINES $450 - But:
- Business trip (reimbursable)
- Personal vacation (not)
- Mix of both?

Context determines category, not the merchant.

More Tricky Scenarios

The Cash Back vs Purchase

; Grocery store with cash back
2026-02-01 * "KROGER" "Groceries + cash back"
  Expenses:Food:Groceries  85.00 USD
  Assets:Cash              40.00 USD
  Assets:Bank:Checking   -125.00 USD

AI sees one transaction, but it is really two.

The Subscription That Changed

; Netflix raised prices but kept the same merchant ID
2026-01-15 * "NETFLIX" "Monthly subscription"
  Expenses:Entertainment:Streaming  15.49 USD  ; was $9.99
  Assets:Bank:Checking             -15.49 USD

Same merchant, different amount - is it an error?

The Medical vs Wellness Gray Area

; Gym membership for physical therapy?
2026-02-01 * "LIFETIME FITNESS" "Monthly membership"
  Expenses:Health:Medical  89.00 USD  ; or Expenses:Personal:Fitness?
  Assets:Bank:Checking    -89.00 USD
  note: "Doctor prescribed for back rehabilitation"

The Human Advantage

What humans bring that AI cannot:

  1. Context - You know why you bought something
  2. Intent - Business or personal purpose
  3. Memory - “That was the trip where…”
  4. Judgment - Gray areas require decisions

Questions

  1. What edge cases have tripped up your categorization?
  2. Do you maintain a list of merchant-to-category mappings?
  3. How do you handle the truly ambiguous transactions?