Every Budget Is a Different Spreadsheet

Anyone applying for film funding submits a budget. As an Excel file, of course. And every production company builds its spreadsheets differently: different sheets, different totals rows, different account names. Until now, the contracts department read every file by hand.

Today a three-stage import pipeline takes over: detect the structure, map the line items, validate the result. In its first test run it correctly ingested 20 of 22 real funding projects, and it maps more than 90 percent of line items automatically.

Import pipeline 22 projects

01 detect structure 22/22

02 map line items >90 %

03 validate result 2 open

no figure leaves the building

Automatically mapped: >90% of line items
First test run: 20 of 22 real projects
Tested against: 45 real files
Sent to the LLM: structure only, not a single number

Deterministic First, LLM Last

The pipeline runs on a clear division of labor. Deterministic matchers go first: exact, alias, fuzzy, phonetic, all aware of the account tree. They handle everything that string metrics can settle. The LLM deliberately comes last and takes on only the semantic remainder no comparison algorithm can read. The industry’s standard budgeting software is supported directly along the way.

Confidentiality is not a setting here, it is architecture: the model receives a masked excerpt of the file in which every number and every formula has been replaced with a placeholder. Only labels and structure ever reach the LLM. Hallucinated accounts are filtered out against the real taxonomy, and a validation layer checks the result against the totals in the source file.

And quality can only go up: every human correction becomes a locked regression test that safeguards future runs.