B2A, short for business-to-agent, describes a company whose immediate customer is a piece of software rather than a person: an artificial intelligence (AI) agent that searches, compares and sometimes pays on someone's behalf. The label borrows its grammar from B2B (business-to-business) and B2C (business-to-consumer). It exists because a growing share of product research, and a small but measurable share of purchasing, now passes through agents built on large language models (LLMs), and those agents read a business differently from the way a person does. They do not see a banner, watch a pre-roll or respond to a hero image. They parse data, call interfaces and apply the constraints their user set.

The term is a framing rather than a standard. No standards body defines it, and it has no specification or version number. It is used by investors, consultancies and vendors to describe a shift in who, or what, a business has to persuade.

How a B2A relationship works

In a B2C sale, a business addresses a person directly. In a B2A sale, a principal - usually a consumer, sometimes a company's procurement team - delegates a task to an agent, and the agent deals with the business. Some writers stretch the label to B2A2C (business-to-agent-to-consumer) to stress that a human still sits at the end of the chain and still carries the money.

A worked example shows the layers. A shopper asks an assistant for trail-running shoes under $150 in a size 10, delivered by Friday. The agent does not scroll a homepage. It queries what it can reach: product feeds, structured data embedded in pages, a merchant's application programming interface (API), or the page itself if nothing better exists. It filters on attributes - price, size availability, delivery date, return terms - and either presents a shortlist or, if its user has granted authority, completes checkout. Every step that a marketer would normally influence with creative, placement or a retargeting impression has been replaced by a machine reading fields.

Companies preparing for that buyer work on four layers.

Readability. Agents need content they can parse cheaply. That means structured product data, consistent identifiers such as Global Trade Item Numbers (GTINs), schema.org markup and clean feeds. Some sites also publish llms.txt, a Markdown index proposed by Jeremy Howard of Answer.AI on September 3, 2024, though Ahrefs found in May 2026 that 97% of such files received no requests at all. Microsoft introduced NLWeb at its Build conference in May 2025 to give sites a natural-language query endpoint for agents.

Actionability. Reading is not buying. Businesses that want agents to transact expose tools. The Model Context Protocol (MCP), published by Anthropic in November 2024, lets an agent call a company's functions; Salesforce's Headless 360, announced on April 15, 2026, exposed its platform to agents as APIs and MCP tools. In the browser, WebMCP, a proposal written by engineers at Microsoft and Google, lets a page register functions such as search_catalog or update_cart, and it reached developers through a Chrome 149 origin trial. Checkout protocols sit on top: OpenAI and Stripe's Agentic Commerce Protocol (ACP) and Google's Universal Commerce Protocol (UCP), launched on January 11, 2026, under which the merchant declares capabilities such as Checkout, Order and Identity Linking in a profile an agent can read.

Identity and access. A business has to decide which agents to let in. Cryptographic schemes such as the Internet Engineering Task Force draft Web Bot Auth have agents sign each HTTP request. Google added a Google-Agent fetcher to its crawler documentation in March 2026, with its own IP ranges. The commercial choice - admit, block, or charge - is made in bot management settings, often by a security team rather than a marketing one.

Payment authority. Card networks issue agent-specific credentials. Mastercard's Agent Pay ties each network token to the agent that made the purchase, and Visa's Intelligent Commerce releases a payment credential only after the purchase is matched to the consumer's instruction.

On the sell side, B2A work involves merchants, commerce platforms such as Shopify and Salesforce, content delivery networks and bot management vendors. On the buy side sit the agent operators - OpenAI, Google, Meta, Perplexity, Anthropic, Amazon - and behind them the people and companies that delegate tasks.

A related design discipline is agent experience (AX), a term promoted by Mathias Biilmann, chief executive of Netlify, in early 2025 to describe how easily an agent can use a product. Microsoft's developer blog published its own explainer of AX on October 6, 2026. AX is the interface layer; B2A is the commercial relationship it serves.

Origin and evolution

The abbreviation is older than the AI meaning. In e-commerce and e-government literature, B2A has long stood for business-to-administration, meaning transactions between companies and public authorities, such as electronic tax filing or procurement portals. Travel technology also uses "B2A" for booking portals built for human travel agents.

No individual or organisation is credited with coining the AI sense. It surfaced in vendor and consultancy writing during 2025, as agentic browsers and shopping agents arrived, and spread through 2026. A March 2026 white paper from the agency 51Blocks, Kantar commentary on "the next retail disruption", a Thoughtworks article on "preparing for the business to agents (B2A) era" and a Fast Company piece headlined "Welcome to B2A, or business-to-agent marketing" all used it. Brand Genetics, a consultancy, expanded the letters differently, as "business-to-algorithm".

The B2B side gained a large forecast. Gartner projected that AI agents would intermediate $15 trillion of business-to-business purchases by 2028, according to Digital Commerce 360's report of November 28, 2025. That figure measures purchases agents touch, not purchases they complete alone.

The term gained its widest audience on November 11, 2026, when Andreessen Horowitz (a16z) partners Olivia Moore and Justine Moore published "Welcome to the B2A Economy". The firm's promotion of the essay cited HUMAN Security data showing AI agent web traffic growing eightfold between August and October 2026, described the market as being in its "early innings", and argued that businesses treating agents as "first-class customers" would capture the growth. PPC Land tested the essay against traffic data the same week. A venture firm that funds agent companies has an interest in the thesis, a point worth weighing alongside the numbers.

Why marketers are paying attention

Most digital advertising is priced on a human seeing something. A B2A transaction can skip that. PPC Land has argued that Amazon's $69 billion advertising business hangs on a browser lawsuit for exactly this reason: an agent shopping inside Amazon need not look at the sponsored listings that fund the marketplace. Coupon and deal sites face a narrower version, since agent-side discounting removes the click affiliates are paid on.

The agent also has its own selection rules, which a business cannot buy its way around with creative. PPC Land reported in October 2026 that Meta's Muse agent was instructed to fill its top five picks with well-known merchants, a rule that decides which sites can receive any referral. For brands, that shifts budget and attention towards product data quality, availability signals, review content and the technical access described above.

Limitations and disputes

Small base. Growth rates are steep because volumes are low. HUMAN's 2025 data showed checkout pages drawing only 2.31% of agent activity against 77% for product and search pages, according to PPC Land's summary. HUMAN sells agent-detection products and does not publish sample sizes.

Reluctant principals. A Salesforce survey of more than 2,500 shoppers in four countries found 76% did not want an agent buying for them. If people keep the final click, B2A describes research more than commerce.

Businesses that refuse. Amazon blocked Meta's Muse on September 20, 2026, and Blend's audit found bot protection refused its shopping agent at 29 of 200 retailers. Many merchants want agents' orders without surrendering the customer relationship, the ad revenue or the data.

Weak readability standards. Google Search said in its May 2026 guidance that it does not use llms.txt, and Google's John Mueller and Microsoft's Fabrice Canel said separate Markdown pages built only for machines are unnecessary. The "agent-ready" checklist sold by agencies remains partly speculative.

Whose agent. An agent owned by a platform that sells advertising may not act purely for its user. No regulator has set a disclosure standard for paid influence on an agent's choices as of November 2026.

Agentic commerce. The market and infrastructure in which agents buy - protocols, tokens, checkouts. B2A is the business-side view of that market: the agent as customer.

GEO. Generative engine optimization aims to get a brand cited in AI-generated answers. B2A is broader, covering transactions, access and payment, not only mentions.

A2A. Agent-to-agent communication, as in Google's Agent2Agent protocol, where two pieces of software negotiate. In a B2A setup, a business's own agent may talk to a customer's agent, but the term describes the commercial relationship, not the protocol.

Business-to-administration. The older meaning of the same letters, for company-to-government transactions.

Recent developments

November 2026 has been crowded. The a16z essay landed on November 11, the same day United Airlines barred unauthorised third-party agents from its bookings, and on November 12 Mastercard said an AI travel agent had "searched, booked, paid for and ticketed a live flight" using an Agent Pay token, according to PYMNTS. The UK Information Commissioner's Office has given agent developers until November 20 to submit evidence for its work on agentic AI. Whether B2A names a lasting channel or a holiday-season talking point should become clearer once Cyber Week traffic is counted.

Timeline

  • Late 1990s onward - "B2A" used for business-to-administration transactions in e-commerce and e-government writing
  • September 3, 2024 - Jeremy Howard publishes the llms.txt proposal
  • November 2024 - Anthropic publishes the Model Context Protocol
  • Early 2025 - Netlify's Mathias Biilmann promotes "agent experience" (AX)
  • April 29-30, 2025 - Mastercard announces Agent Pay; Visa announces Intelligent Commerce
  • May 2025 - Microsoft introduces NLWeb at Build
  • September 29, 2025 - OpenAI and Stripe release the Agentic Commerce Protocol
  • November 28, 2025 - Gartner forecast of $15 trillion in agent-intermediated B2B purchases by 2028 reported
  • January 11, 2026 - Google launches the Universal Commerce Protocol
  • March 2026 - Google adds Google-Agent to its crawler documentation; 51Blocks publishes a B2A marketing white paper
  • April 15, 2026 - Salesforce announces Headless 360
  • September 20, 2026 - Amazon blocks Meta's Muse
  • October 6, 2026 - Microsoft's developer blog publishes an AX explainer
  • November 11, 2026 - a16z publishes "Welcome to the B2A Economy"; United Airlines bars unauthorised AI agents
  • November 12, 2026 - Mastercard reports a live flight booked and paid for by an AI agent

Summary

Who. Businesses selling to AI agents - retailers, airlines, software vendors and the platforms that serve them, such as Shopify, Salesforce and Cloudflare - and agent operators including OpenAI, Google, Meta, Perplexity, Anthropic and Amazon. Investors such as a16z, consultancies such as Kantar and Thoughtworks, and Gartner have popularised the framing.

What. A commercial relationship in which an AI agent, acting for a consumer or company, is the party a business must reach and persuade, through readable data, callable tools, verifiable identity and agent-specific payment credentials rather than through advertising aimed at human attention.

When. The AI sense spread from 2025, built on infrastructure dating from late 2024, and reached mainstream attention with a16z's essay of November 11, 2026. The same letters have older meanings in e-government and travel.

Where. On merchant websites, APIs and MCP servers, inside chat assistants, agentic browsers and messaging apps, and in B2B procurement, with most consumer activity in the United States as of November 2026.

Why. Agents strip out the human viewing moment that most advertising is priced on, so businesses that want agents' orders have to compete on data, access and price, while low volumes, consumer reluctance, merchant blocking and unsettled rules keep the scale of B2A in dispute.