AI labels, code counts
How we split work between AI and plain code when reading documents, and why we never let the model do arithmetic.
· 2 min read · For business owners buying AI software
A lot of what our products do is read: answer sheets, news stories, bank statements, long PDFs. Somewhere in each is the one fact that matters. The interesting part is how we split the work between AI and plain code.
The split
- Code finds candidates. Simple rules pull out the lines that could matter: amounts, dates, place names.
- AI labels them. A small model decides what each line means. Is this a debit or a refund? Is this story about a road accident in Madurai?
- Code does the numbers. Amounts and dates are read with plain, tested parsing. Totals, comparisons and alerts are ordinary code.
Models are good at judging what a sentence means. They are not a calculator.
In Kanakku, AI reads a bank statement line and suggests which ledger it belongs to. The amount itself is parsed by code and posted by double-entry rules, so the books always balance. There is no chance of the model "rounding" a number.
NewsDesk works the same way
AI tags a news story by district, department and severity, and writes a two-line summary. Code does the counting, sorting and alerting. If a tag is wrong, one story is misfiled. If a model did the counting, every report would be suspect.
When a person should decide
Some steps should not be automated at all. When a website asks a human to prove they are human, a human answers. When a mark goes on a student's record, a teacher confirms it. It is a boring split, and boring is what you want in production.
