It Says Director, It Means Directing

To know whether one production cost more than another, you have to be able to lay their budgets side by side. That is where the trouble starts: every production company structures its budget by its own logic. The same service carries a technical term in one budget, a job title in another, and vanishes into a catch-all line in a third. Before anything can be compared, every single line has to be translated onto a common chart of accounts.

Classic text matching gets you part of the way — and then no further. Roughly 30 percent of the line items can be matched automatically through names, spelling variants, and similarity measures. The rest fails not on typos but on meaning. That is why a language model sits at the end of a chain of deterministic methods: it only gets the cases none of them could solve.

Chain of methods

exact match 12 %

spelling variants 11 %

similarity score 7 %

the rest needs meaning

Language model 70 %

String matching alone: ~30% of line items
Order of operations: Deterministic first, then the LLM
Full disclosure: the gain hasn't been measured yet

Where Text Matching Goes Blind

One example: a budget line says “director” — what it means is the directing account. Vehicle costs, meanwhile, turn up under transport, under production design, or with the crew, depending on the production. A person assigns lines like these without a second thought. An algorithm comparing character strings sees only words that bear no resemblance to each other.

These are exactly the cases the language model takes on: the remainder the deterministic chain leaves behind, the lines that only yield to meaning. A side benefit falls out along the way, one nobody had captured systematically before — when personal names appear in the free-text fields, the model spots them and matches them against the master data.

That leaves the question of how much all this is worth. The honest answer: measurement has only just begun. By how many percentage points the language model actually raises the automatic match rate is not yet settled. What every run already shows: it hits exactly the lines where every other method fails.