Appendix C: From B2Ai to B2A — Agentic Commerce and the Selection ≠ Transaction Gap — Ernest D. Johnson — ToastDeck Research

C.1   Thesis Extension: B2Ai Is the Foundation; B2A Is the Execution Horizon

The central thesis of the paper — Recognition ≠ Selection — addresses the upstream layer where AI systems interpret, represent, compare, trust, and select businesses before presenting options to humans or downstream workflows. That layer remains necessary. But the market is moving from AI as an answer surface toward AI as an execution surface.

This transition can be described as the movement from B2Ai to B2A: from businesses being interpreted and selected by AI systems to businesses being discovered, invoked, delegated to, transacted with, and reconciled by AI agents.

C.2   The Protocol Layer: The Commercial Internet Is Gaining Machine-Native Execution Rails

The emerging B2A stack is not one protocol. It is a protocol environment. UCP, AP2, A2A, MCP, ACP, and x402 address different boundaries: commerce flows, payment authorization, agent-to-agent coordination, tool and data access, conversational checkout, and HTTP-native payments.

• UCP — agentic commerce. Google describes UCP as an open-source standard for agentic commerce providing a common language and functional primitives for consumer surfaces, businesses, and payment providers. Supports discovery, capability profiles, checkout, order management, and AP2 compatibility.¹
• AP2 — agent payments. Google describes AP2 as an open protocol for secure agent-led payments across platforms, with a need to authenticate, validate, and convey an agent’s authority to transact.²
• A2A — agent-to-agent coordination. Google launched A2A in April 2025 so agents built by different vendors can communicate, exchange information securely, and coordinate actions. The Linux Foundation later hosted the project, reporting 150+ supporting organizations.^{3 4 5}
• MCP — agent-to-tool/data access. Anthropic describes MCP as an open standard for secure two-way connections between data sources and AI-powered tools.⁶
• ACP — conversational commerce checkout. Stripe says ACP is an open standard co-developed with OpenAI to let AI agents and businesses complete purchases while merchants remain merchant of record.⁷
• x402 — HTTP-native payment. Coinbase describes x402 as a standard for internet-native payments using the 402 Payment Required pattern.⁸

The important point for B2Ai is not which protocol wins. The commercial internet is gaining machine-native execution rails, and the protocols are converging rather than fragmenting. Once agents can act through standardized, interlocking protocols, the upstream selection question becomes commercially sharper: why this business instead of another?

C.3   The Payment and Trust Layer

Agentic commerce requires more than checkout. It requires identity, permission, spend control, authentication, fraud management, settlement, auditability, and dispute handling.

• Visa Intelligent Commerce is positioned around AI-agent commerce, including tokenized payments, authentication APIs, and trusted-agent controls.⁹
• Mastercard Agent Pay introduces agentic payments and Agentic Tokens built on existing tokenization capabilities; Mastercard describes trusted agents being registered and verified before making secure payments.¹⁰
• Stripe and OpenAI’s ACP powers Instant Checkout in ChatGPT and defines a merchant-friendly open standard for agentic commerce.⁷
• PayPal launched agentic commerce services including agent-ready payments and Store Sync, designed to make merchant product data discoverable and purchasable in AI-driven shopping experiences.¹¹
• Skyfire provides verified agent identity and payment credentials, including tokenized cards, micropayments, and programmatic checkout.¹²
• Nevermined positions itself as payments infrastructure for AI agents, APIs, datasets, and digital services, with support across MCP, x402, A2A, and AP2-oriented workflows.¹³
• Payman focuses on agentic banking: AI agents executing real banking transactions on existing rails.¹⁴

This layer matters because an agent cannot simply “buy.” It must prove authorization, operate within policy, settle payment, support reconciliation, and leave an audit trail. Businesses that cannot participate in this layer will be selected and then fail the transaction.

C.4   The Execution and Orchestration Layer

• Arcade is building an MCP runtime for production AI agents, emphasizing secure agent authorization, reliable tools, governance, and user-specific permissions.¹⁵
• Browserbase is building browser infrastructure for agents, including Search API, Fetch API, and Browser-as-a-Service so agents can navigate, log in, and perform tasks where APIs do not exist.¹⁶
• Merge Agent Handler connects AI agents to third-party applications, manages authentication and permissions, and provides evaluations for tool reliability.¹⁷
• Alibaba’s Accio Work is an agentic AI platform for SMEs that automates complex business operations with permissions for high-stakes tasks including financial transactions.¹⁸
• Anthropic’s Project Deal should be treated as an experiment rather than production infrastructure. Run in December 2025 and published April 2026, it placed Claude agents in an internal classified marketplace where they autonomously negotiated and closed 186 real transactions worth over $4,000 on behalf of 69 employees, with no human intervention in the dealmaking itself. The experiment also documented a capability asymmetry: parties represented by a stronger model systematically obtained better outcomes — an early empirical signal that agent quality itself becomes a selection-adjacent commercial variable.¹⁹

C.5   Why B2Ai Still Comes First

B2A does not eliminate the need for B2Ai. It makes B2Ai more important.

Before an agent can transact with a business, the business must first enter the agent’s candidate environment. It must be reachable, parseable, discoverable, recognized, represented, trusted, and selected. UCP can standardize commerce flows. AP2 can standardize agent payment authorization. A2A can standardize agent coordination. MCP can standardize tool access. ACP can standardize conversational checkout. x402 can standardize HTTP-native payments. None of those protocols answer the upstream question: why this business?

That remains the B2Ai selection problem.

This extension also sits on the accountability axis established in Section 5 of the main thesis. Where Section 5 holds that a model substitutes only when it can become the accountable party, B2A describes the complementary case: the agent acts, but an accountable business must still be selected. Under ACP, for example, the merchant remains merchant of record even when the agent completes the purchase. Selection therefore remains the upstream event, and the transaction inherits its quality.

C.6   How B2A Extends the Eight Layers

• Layer 0 expands from crawlability into capability access: can an agent discover, invoke, and safely interact with the business’s machine-readable surfaces? Agent identity verification becomes the agentic evolution of the 0A Access condition.
• Layers 1–3 expand from identity clarity into operational entity consistency across websites, APIs, catalogs, feeds, merchant profiles, payment credentials, service regions, and third-party sources.
• Layer 4 expands from authority signals into transactional trust: licensing, policies, identity verification, fraud controls, payment tokenization, SLAs, return terms, liability limits, and permissioned access.
• Layer 5 becomes more consequential because selection may trigger action: cart building, checkout, booking, scheduling, negotiation, routing, or task delegation.
• Layer 6 becomes an auditability layer: why was the entity selected, what evidence supported the action, what authorization existed, and what constraints governed the transaction?
• Layer 7 becomes operational freshness: stale inventory, price, availability, credentials, service boundaries, or legal terms can now break a transaction, not merely weaken a recommendation.

C.7   Working Definition

C.8   What Businesses Will Need to Expose

A B2A-ready business will need more than a crawlable website. It will need an execution surface. The specific requirements will vary by industry, but the categories are already visible:

• Machine-readable capability descriptions: what the business can do, for whom, under what constraints.
• Real-time operational state: pricing, inventory, availability, appointment windows, location, fulfillment rules, and status.
• Authorization and payment readiness: accepted payment methods, agent-payment support, spend limits, identity verification, merchant-of-record rules, and dispute and return terms.
• Agent identity handling: which agents are allowed to act, what credentials they present, and how the business distinguishes legitimate agents from malicious automation.
• Audit and reconciliation: logs of the agent’s request, user intent, authorization state, action taken, payment status, fulfillment status, and exceptions.
• Fallback and human handoff: when the agent must ask for approval, defer to a human, or stop the transaction because risk or ambiguity exceeds the policy boundary.

C.9   Outlook

As of mid-2026, the market appears to be in a transitional phase. Many agentic commerce flows still include human approval, merchant onboarding, sandboxing, limited partner rollout, or experimental conditions. But the infrastructure direction is clear: protocols for agent communication, tool access, commerce actions, and payment authorization are being standardized across major platforms, payment networks, and specialized infrastructure startups.

That matters for businesses. The next competitive boundary will not be only who appears in AI answers. It will be who can be selected by AI systems and then safely acted upon by AI agents.

Source Verification Matrix