Requirement evaluated: The AI coding model must learn from this buyer's specific transaction history to improve dimension coding accuracy over time, using the 12,000 monthly invoices as the training corpus. The vendor must explain the actual mechanism (per-customer model, fine-tuning on approval history, rules derived from prior accepted coding, or equivalent) and must not describe a generic pretrained model as if it were customer-specific learning. The buyer's question, 'how does the per-customer model learn from our history,' must be answerable with a concrete mechanism and a measurable lift curve, not a marketing claim.
For a buyer processing 12,000 invoices a month across dozens of NetSuite coding fields, Medius's learning mechanism is delivered through SmartFlow, a CNN-based proprietary model that auto-codes GL account, tax fields, approver values, and coding dimensions for non-PO invoices. The mechanism is explicitly company-specific: according to Medius's invoice automation product page, SmartFlow is 'trained on your historical actions and enriched by 2.4 billion+ invoice field data points across Medius's global customer base,' and a Medius Chief Architect confirmed in a published interview that 'our machine learning technology uses pattern recognition to capture invoices, code them correctly, and route …
Limitations: Medius does not publish a measurable lift curve showing accuracy improvement as a function of invoice volume past the cold-start '95% after two invoices' benchmark, so the buyer cannot verify the concrete progression their 12,000 monthly invoices would drive. …