Commerce-Native AI: Why Transformers Are Fundamentally Wrong for Commerce
Transformers were built for language. Commerce needs its own architecture — with formal verification, state machines, and learnable market physics.
Transformer architectures were built for language. We applied them to commerce — and wonder why AI agents hallucinate, overshoot budgets, and violate policies. This article sketches an alternative approach: Commerce-Native AI (CVAI). The Problem: Commerce Is Not a Language Problem Everyone building an AI agent for commerce today uses a language model as foundation. This creates a fundamental mismatch: What Language Needs What Commerce Needs Sequential token prediction Multi-temporal state management Probabilistic truth Binary truth (deal or no deal) No persistent state State IS the core (inventory, budget, intent) Single speaker N simultaneous parties with private interests Hallucinations tolerable Zero tolerance: errors = money lost We are not proposing better prompts. We are proposing a different architecture. Three Paradigm Shifts 1. Transactions Are Geometric Objects Inspired by recent developments in 3D reconstruction, we hypothesize: A commerce transaction exists in a latent space. It has structure (price, quantity, terms) and context (trust, market conditions, history). From a single intent signal, you can reconstruct the entire transaction dynamics — buyer, seller, and platform perspectives. This is not a metaphor. This is a trainable architecture. 2. Agents Must Prove, Not Promise Formal verification — known from mathematics (Lean 4, Coq) — is increasingly applied to AI outputs. Projects like Leanstral show that models with 6 billion parameters can generate formally correct proofs. Applied to commerce: Budget proof: The agent proves its bid does not exceed the budget Policy proof: The agent proves its action complies with all policies Double-spend proof: The agent proves no amount is spent twice No proof chain → no transaction. 3. Markets Have Physics Prices and demand follow laws. Trust builds slowly and breaks fast. Network effects follow power laws. These regularities are learnable. A Commerce World Model learns this physics from data. It can simulate transactions before execution, predict market movements, and detect anomalies. Crucially, commerce physics is low-dimensional — such a model could run on consumer hardware. The CVAI Architecture Four modules form the core: Commerce State Machine (C-SSM): A state-space model (not transformer) for transaction states. Linear complexity O(n) instead of quadratic. Commerce Latent Space (CLS): Transactions as geometric objects. Multi-view encoding allows reconstructing any stakeholder perspective from a single signal. Proof-Generating Agent (PGA): Every agent output comes with a formal proof. No proof, no action. Commerce World Model (CWM): Learns market physics. Enables simulation before execution. Open Research Questions This is not a finished system. …
- Author
- Holger von Ellerts
- Published
- 2026-03-19
- Topics
- Research, CVAI, Agentic