Six vendors. 1,200 commercial invoices. One human-reviewed gold set. This is the benchmark we wish had existed before we built Floowed's extraction engine.
What we measured
Accuracy (line-item match), median extract latency, P95 latency, cost per document, and failure mode distribution.
Where Floowed sits
Median extract time: 3.2 seconds. P95: 8.1s. Line-item accuracy: 97.3% on standard invoices. Cost at scale: $0.004 per document.
The five failure modes
Handwritten annotations, multi-currency line items, non-standard date formats, merged cells, and vendor abbreviations. Every vendor hits all five. We handle the first three better than average. The same gap shows up on bank statements, which is why frontier AI can't read bank statements without a purpose-built layer, and it is the core of document extraction accuracy in AI automation.
Floowed's AP and BPO document automation handles three-way match, GL coding, and ERP posting with preset flows your team configures and runs directly.