That’s fair — and I agree if enough context exists.

The key limitation is that a raw bank transaction usually contains very little semantic information: amount, merchant name, date. From that alone, an LLM can only guess based on patterns or prior behavior, not actually know what the expense was for.

“$100 at a supermarket” could be groceries, pet food, a household item, or something work-related. An LLM can infer probabilities once it has enough historical data and feedback, but that still means the user has to correct or confirm things at some point.

So I see LLMs as very helpful for assisting categorization (suggestions, defaults, learning over time), but they can’t fully replace intent unless the underlying data becomes richer than what bank statements provide today.