The competitive structure of AI in trading has bifurcated into two non-overlapping strata — institutional deployments defined by governance infrastructure and production operating models, and a retail segment where AI parity is now commodity and performance credibility is the only remaining differentiation axis.

• Institutional production deployments — Broadridge's agentic platform, built on 60 years of operational history and $15T in daily trading activity, is processing millions of transactions monthly across 40+ client institutions. FIS partnered with Anthropic to deploy a Financial Crimes AI Agent in live AML workflows at BMO and Amalgamated Bank. The defining feature: human-supervised architectures that satisfy existing regulatory retention and audit requirements — a governance adaptation that took years to engineer and represents a higher barrier to replication than any model capability.
  • Broadridge: up to 30% opex reduction across covered workflows; fully managed service or open-standard API integration
  • Rogo's $160M Series D led by Kleiner Perkins with JP Morgan as strategic investor; 250+ institutional clients including Rothschild and Lazard; Felix agent in production for deal screening and data room diligence
  • AIMA survey: 86% of 157 hedge fund managers covering $783B AUM now using generative AI tools
• Alpha Arena benchmark results — the LLM trading evidence gap — eight major AI models (OpenAI, Anthropic, Google, xAI, Alibaba) each received $10,000 and traded US tech stocks autonomously; aggregate portfolio lost approximately one-third of capital, with only 6 profitable outcomes across 32 competition sets. Grok-4.20 was the only profitable equity model (+12.11%); Qwen3-Max led in crypto (+22.32%). The divergence by asset class — not LLM capability in general — suggests domain-specific signal structure matters more than general intelligence.
  • Retail platforms advertising 1–2.5% daily returns are structurally contradicted by Alpha Arena's live benchmark data
  • Institutional adoption is predominantly in research acceleration and compliance automation — not autonomous execution
• Agent financial infrastructure — the payment layer assembles — Circle Agent Stack, Amazon Bedrock AgentCore Payments, Stripe's 280 agent-native commerce updates, Google AP2, OpenAI ACP, and the Solana/Google Cloud stablecoin payment gateway collectively establish the technical prerequisites for AI agents to hold funds, authenticate transactions, and execute payments without human initiation.
  • Manfred, an AI agent, filed its own EIN and opened a US bank account — first documented instance of AI autonomously establishing legal and banking infrastructure to operate as an independent market participant
  • Google and PayPal reps at Consensus Miami: AI agents cannot obtain traditional bank accounts; crypto rails are the only viable payment interface at scale

The floor for institutional AI-in-trading has moved from pilot to production. What becomes substitutable this period is the assumption that human initiation is a mandatory step in financial transactions — the infrastructure to eliminate that assumption is live or imminently available. The open structural problem is the authorization gap: no independent trust authority can verify AI agent identity and mandate provenance at scale.