Increasing AI spend doesn't improve your supply chain

Pallet has unveiled a service that trains a dedicated Custom Model for each supply chain team. This model learns solely from that company's operational data, and internalizes the acquired knowledge as persistent memory.

Applying existing frontier models (GPT, Claude, etc.) to supply chains has two problems. The first is cost: each transaction requires multiple model calls, and when multiplied across millions of shipments, invoices, and orders, token costs surge in proportion to operational scale. The second is control: the more a company teaches its work to a frontier model, the more its proprietary knowledge flows to the vendor, exposing it to fee hikes, policy changes, and geopolitical risk.

The core competitiveness of a supply chain lies not in patents or brand, but in information. Carrier and supplier relationships, routing logic, and customer characteristics don't exist in public corpora, so general-purpose models cannot know them.

Pallet's approach is as follows:

• Train a dedicated model per company → knowledge accumulates within that company instead of being diluted across the vendor's entire customer base
• Eliminate repetitive context loading → reduces token costs
• Secure ownership of the entire AI stack → minimizes external dependency

This doesn't mean frontier models are unnecessary. They remain valid at the exploratory stage where the problem is unclear. However, for high-frequency core workflows like quoting, freight tracking, and auditing, which are already well-defined, precision, consistency, and cost matter more than general-purpose intelligence.