AI extraction with a human in the loop, five document types

What gets extracted

LS Technologies, an electronics-components import/export trading business, runs five kinds of documents through AI extraction before anything reaches Tally: purchase invoices, freight invoices, bills of entry, purchase orders, and bank statements. Each has its own prompt tailored to that document's fields — a purchase invoice and a customs bill of entry don't share a structure, and treating them as one generic "document" would lose accuracy on both.

How it reads a PDF

There's no separate OCR step. The PDF is sent straight to the model — an LLM with a file-parsing capability reads it directly, rather than running text extraction first and handing the model plain text. Bank statements get special handling: because they can run to many pages, the document is processed in chunks and the results are stitched back together and de-duplicated, rather than trying to reason over one long document in a single pass.

How accurate it actually is

Reviewed against real usage, here's the share of each document type that needs zero correction before it's approved:

• Purchase order: 95%
• Purchase invoice: 94%
• Bank statement: 89%
• Freight invoice: 91%
• Bill of entry: 88%

Bills of entry come in lowest — customs documentation is the least standardized of the five, which tracks with what you'd expect from the least uniform paperwork. Every number here still means a meaningful share of documents need a correction, which is exactly why extraction alone was never the plan.

How do these numbers compare with advertised accuracy?

Invoice-automation tools commonly advertise 95–99% accuracy. A 2026 benchmark review found those claims have converged in that range while being measured under conditions each vendor chooses, with no independent benchmark behind them (Parseur). The figures above are lower, and deliberately so: they're per document type, from real usage, and they count any document a person had to correct — including customs paperwork such as bills of entry. A lower number measured honestly is more useful for planning than a higher one measured on a vendor's own samples.

What happens to the rest

Nothing gets extracted and posted in one step. Every document — regardless of how confident the extraction looked — sits in a review state until a person checks it and approves it. That's not a hedge against a bad model; it's the same decision this practice makes anywhere a model's output could turn into a financial record: draft with AI, confirm with a person, post only after that.

purchase invoice zero-correction rate

94%

freight invoice zero-correction rate

91%

bill of entry zero-correction rate

88%

purchase order zero-correction rate

95%

bank statement zero-correction rate

89%