AI Data Mapping

A regional film fund receives production cost calculations in every conceivable Excel format – exports from the industry's standard calculation software, hand-built sheets, bilingual layouts. The import pipeline works in three stages: an LLM reads only the layout of a masked digest and returns a machine-executable structure spec, deterministic code applies it to the real cells, and mapping onto the industry-standard account tree is done by recommenders – deterministic first, LLM last. In real-world testing, 20 of 22 projects were imported, with over 90% of positions auto-mapped.

A major studio's budgeting system normalises incoming production budgets – each in its own account structure, language, and naming habits – onto a target account tree. The deterministic matcher chain plateaued at roughly a third of the lines; the LLM runs deliberately last and reads what string metrics cannot: that a director's role label belongs on the directing account, or that a vehicle line is a transport cost. Confident suggestions auto-apply through the same threshold machinery as all matchers – the rest surface as ranked suggestions for humans.

A licensing and royalty platform for a global media group resolves every incoming title – a distribution partner's spreadsheet row, a localized film title, an episode reference – to the house product identifier. The deterministic mapping engine is exposed as an MCP server: the user's AI client supplies the inputs, the engine passes the verdict – per row an identifier with explicit certainty, or an honest no-match. Answers cannot drift between channels because the same engine decides everywhere.

The same paid LLM call is mined a second time: person names embedded in budget lines are extracted and resolved against a talent database – as a free side effect of the mapping. This replaced a dedicated cloud NER service.