By the numbers: Agent or agentic talk in over 180 sessions. Speakers described agents already working for analysts, traders and contact center teams, then asked what it takes to scale them.

Multiple sessions explored:

• Agents with a wallet: Agents connect discovery, comparison and checkout, and Mastercard expects 1 in 10 transactions to be agent-led by 2030, or sooner
• Banks at scale: BNY put agents in front of every employee and measured only adoption through 2024 and 2025
• Support turns profit center: Customer service shifts from a cost center to a revenue line
• Legacy rewritten for banks: AI maintains or rebuilds old bank systems, a point raised in 28 sessions

That raised harder questions about:

• Pilot to production (66 sessions): Smaller teams and faster builds don't guarantee better margins once compute and operating costs are counted
• Data first: Agents need consistent, authoritative data to make sound decisions
• Permission (58 sessions): Spending limits, merchant restrictions and human approvals have to be built into the payment architecture
• Proof of trust: Identity, customer intent and authorization must travel with the transaction, and "Know Your Agent" came up in 13 sessions
• Liability (52 sessions): Someone must own disputes, compromised agents and unauthorized purchases
• Guardrails (45 sessions): Human-in-the-loop controls

Permission was the most contested question. One panel cited research in which 3 in 4 US AI users wanted an agent to act, not just advise. Speakers argued that consent frameworks built for people need policy, budget and timing limits for agents, and that many so-called agentic payments are really a chat triggering an old workflow. Jefferies still doesn't let colleagues share agents, because access control isn't solved.

The broader implication: the model is no longer the bottleneck. Data, access control, cost and accountability are.