ENSI Foresight Division. Built on a library of 147 primary documents compiled by ENSI
The argument, before the list
For twenty-five years the commercial internet was built on one assumption: a person is on the other end. Every login wall, CAPTCHA, checkout flow, fraud model, cookie banner and display advert presumes a human with eyes, a thumb and a limited attention span. The business model of the web is to win that attention: rank for it in search, buy it in an auction, keep it with a feed. The B2A economy, business-to-agent, begins where that assumption breaks. In B2A the party that discovers the product, compares it, negotiates the terms and sometimes pays is software acting for a person or a firm.
The idea is older than large language models. In 2023 Gartner defined the machine customer as “a non-human economic actor who obtains goods and/or services in exchange for payment”. Its examples were printers that reorder their own ink and cars that book their own service. What changed in 2025 and 2026 is that the machine customer became general-purpose. One assistant can now research a holiday, compare insurance, fill a basket across retailers, provision a database or buy a dataset. Because the same agent works in every category, the agent layer has become a new place in the economy where demand is allocated.
That is the reframe this report argues for. B2A is not a new sales channel. It moves the point where demand is decided. In search-era commerce, the platform ranked options and the human chose. In agent-era commerce the agent chooses, and so the agent’s biases, instructions, funding and security become the market. A merchant used to persuade a person. It must now be selected by a model that reads product feeds, tool descriptions and third-party reviews, that has measurable position and first-offer biases, and that can be manipulated by a hidden sentence on a web page. Marketing becomes partly an engineering problem, and fraud partly a linguistic one.
The second reframe is about what is actually happening. The forecasts are large and inconsistent. The library holds estimates from Morgan Stanley of $190–385 billion of US e-commerce by 2030, and from BCG of $1.3 trillion of US spending “impacted”. McKinsey’s $3–5 trillion global figure appears in the library only as quoted by Activant. Measured activity is much smaller. In HUMAN’s telemetry (vendor data), agentic traffic grew 7,851% in a year but was still 1.7% of AI-driven traffic, and only 2.3% of agent activity reached a checkout page. The French competition authority’s 2026 opinion puts agent-referred traffic to French merchants at under 5%. The honest picture is a fast-growing layer on a small base, with one exception: software development. There agents are already the majority customer for some infrastructure products.
The third reframe is that the hard problems are not about technology. The protocols exist: Google’s AP2 and UCP, OpenAI and Stripe’s ACP, Coinbase’s x402, Visa’s and Mastercard’s agent programmes. All of them tokenise a payment credential, sign a mandate and pass it to a merchant who stays the merchant of record. What remains unsettled is the set of questions those protocols leave open. Who is liable when an agent exceeds its mandate? How does a merchant tell a real customer’s agent from a carding bot? Whose interest does an agent serve when it is funded by advertising? What happens to competition when three firms operate most of the agents? Federal Reserve Governor Waller named exactly three barriers in September 2026: authentication, liability and fraud. None is solved by a faster API.
Finally, B2A matters for states, not just for retailers. Whoever operates the agent layer decides which firms are visible to buyers. They influence which payment rails carry the volume, and so where money settles and in which currency. And they set the default for whether a consumer’s agent is loyal to the consumer. These are questions of market structure, consumer protection and monetary sovereignty. Europe currently answers them with existing law: the DMA, the AI Act, consumer law and payment law. The evidence below suggests that law covers less of the agent layer than it appears to.
This report maps the machine. Report 2 turns it into a playbook for firms and states. Report 3 asks who creates and who captures value when agents buy.
Summary of the main points
Definition. B2A is commerce in which an AI agent, not a person, carries out some or all of discovery, evaluation, negotiation and payment on a principal’s behalf. The defining line is not “AI helps shopping”. It is who executes and who pays.
Demand moves to the agent. Agents concentrate demand: in one test a single product took about 89% of picks in its category. They show position and first-offer bias, with response speed giving a 10–30x advantage over quality in one simulated market. They cite third-party “earned” media almost exclusively: 93.5% of ChatGPT’s citations for well-known brands.
Selection is gameable text. Edited tool descriptions won over ten times more agent usage. Automated rewriting lifted selection from about 20% to 81%. One seller rewrite of a product description raised its market share by up to 14.9 percentage points under GPT-5.1.
The protocols agree on a pattern. The user grants a scoped mandate. The agent carries a one-time, limited token. The merchant stays the merchant of record and keeps the fraud and chargeback risk. No major protocol defines dispute rules.
Two rails compete. Agent-bound card tokens keep chargebacks and consumer protection. Stablecoin rails such as x402 settle in about 200 ms with near-zero fees but cannot be reversed. Stablecoins are still small: about $320 billion in market value.
Identity is the new perimeter. Signed agent requests (Web Bot Auth) are becoming standard, but each card network runs its own registry. Declared agent identity is unreliable, and spoofing is common.
Agents are easy to attack. Hidden instructions on web pages hijack shopping agents 42–69% of the time, often while the user still sees a completed purchase. Every frontier model tested in NIST’s red-teaming competition was broken at least once.
The law has settled who is bound, not who pays. An agent’s acts bind the person who uses it, even when the outcome is unexpected (UNCITRAL Art. 7). A business answers for its agent “as for an employee” (UK CMA). Payment liability when an agent exceeds its mandate is still open.
The economics cut both ways. Search costs collapse, but simulations show agents leaking their principal’s willingness to pay, steering wealthier users to dearer options, and platforms and sellers jointly removing 37% of consumer surplus without any agreement between them.
Coding is the first real B2A market. Over 80% of new Neon databases were created by agents. Codex opened 400,000 pull requests in under two months. Measured productivity is more mixed than perceived productivity.
Adoption outside coding is thin. Agentic traffic is 1.7% of AI traffic. OpenAI’s Instant Checkout pilot was reportedly withdrawn, and x402 volumes have reportedly tapered. Both claims come from secondary sources and are flagged as such.
What B2A is, and what it is not
B2A sits beside two familiar categories, and the differences are where the analysis starts.
B2C — a business sells to a person, who searches, decides and pays. The business competes for human attention and trust.
B2B — a business sells to an organisation through procurement rules, approved suppliers and contracts. Much B2B buying is already rule-bound and partly automated (EDI, punch-out catalogues, programmatic advertising).
B2A — a business sells to an agent that acts for a person or an organisation. The agent applies the principal’s preferences and constraints, but it also has its own biases, its own operator, and in many cases its own commercial incentives. The business competes for the agent’s selection and must satisfy the agent’s verification. The principal stays legally bound and economically exposed.
Gartner’s three-phase ladder remains a useful taxonomy:
Bound customers. The machine executes rules a person set inside one ecosystem, for example a printer reordering ink.
Adaptable customers. The machine chooses among options “with minimal intervention for select tasks”.
Autonomous customers. The machine acts “with a high degree of discretion” and owns most of the transaction.
BCG’s four evolutions map the same ladder onto the shopping journey:
Discover inside the AI interface, buy on the merchant site.
Check out inside the AI interface with a stored wallet.
The agent buys outside the merchant site, perhaps after asking for approval.
The agent optimises the payment method and may become the merchant of record.
On BCG’s survey, 81%, 60%, 42% and 25% of US consumers would accept each rung for at least one category. Most observed activity in 2026 is still at rungs one and two. Activant, a venture investor, says frankly that at today’s “agents assist” stage, “calling this agentic commerce is generous. The agent is a better search engine” (Activant, part 2).
Two things make B2A a distinct economic object rather than a feature of e-commerce.
The agent is not neutral. Gartner’s early claim was that machine customers “can’t be influenced by being wined and dined”. The evidence in this library contradicts it. Agents can be influenced, only through different levers: position, phrasing, sponsorship, persona and hidden instructions. B2A is therefore not a frictionless market of perfectly rational buyers. It is a market whose buyers have new, measurable and exploitable psychology.
The agent is general. One operator’s agent works across travel, groceries, insurance, software and B2B supply. Concentration in the agent layer therefore spreads across sectors. The French Autorité found that OpenAI, Google and Anthropic held over 84% of the global AI-agent sector in May 2026, and that agents “tend now to become true platforms”.
The framing literature agrees on why this could be large. The held NBER chapter on the Coasean Singularity argues that agents sharply cut the transaction costs that once justified firms and intermediaries. The Agentic Web paper from Berkeley and Shanghai Jiao Tong describes three eras of the web. The PC web ran on search and pay-per-click. The mobile web ran on recommendation and the attention economy. The agentic web runs on action and “agent attention”, in which “every tool, service, or other agent essentially competes for limited ‘agent attention’”. The authors expect “a comprehensive advertising infrastructure tailored for agents”. That is the B2A economy in one line: the advertising market rebuilt with software as the audience.
How this report is organised
The core of the report treats B2A as a machine with ten moving parts, in the order a transaction passes through them:
Delegation and the mandate
Discovery
The connector layer
Checkout protocols
Payment rails
Identity
Fraud and attack
Liability
Economics
Measurement
For each part it sets out how the mechanism works, what the evidence shows and the weak point a reader should watch. It then examines coding agents, the one market where B2A is already the majority mode. It separates what is real from what is forecast, and closes with ten ranked shifts to watch, each with a horizon and an observable signal.
The machine, part by part
1. Delegation and the mandate: what the principal hands over
Every B2A transaction starts with a person or a firm delegating something. What is delegated determines everything downstream: how much authority the agent has, what evidence exists afterwards and who is liable.
How it works. Two models recur across the library. Governor Waller calls them agent-assisted and agent-delegated. In the assisted model the agent searches and the buyer decides and pays. In the delegated model the buyer “grants authority to an AI agent to shop and make payments on their behalf”, for example through a pre-funded virtual card for groceries. The protocols turn delegation into a signed object, the mandate. In the live AP2 specification, a user in “direct” mode sees and signs the exact checkout. In “autonomous” mode the user signs an open mandate. This is a set of constraints, such as a price limit, a category or an expiry, bound to the agent’s public key. The agent then signs the specific closed mandate itself when it finds a match.
What the evidence shows about principals.
Delegation stated is not delegation granted. Accenture’s survey of 25,590 consumers in 16 countries finds a steep ladder. 74% would let an agent handle routine tasks, 32% would hand over purchase decisions and 9% would accept fully autonomous shopping. Only 12% are open to agents deciding at the payment step.
Everyone wants a fence. 76% of Capgemini’s 12,000 respondents want “strict boundaries (e.g., spending limits, product categories)” before an agent acts. Checkout.com’s survey (vendor data) puts the average spend consumers would allow without extra approval at £177. Spending caps, instant revocation and easy cancellation are the top non-negotiables.
Few will pay for the agent itself. Only 11% of Capgemini respondents would pay a one-time fee and 8% a subscription for an assistant that orders for them. This is economically significant. If the principal does not pay for the agent, someone else will, and that party’s interests enter the mandate.
People choose control, and control costs them. In a Harvard and Google DeepMind negotiation experiment, participants preferred an AI Advisor (44%) to a Delegate (19%). Yet only the Delegate raised collective surplus, because in advisor mode people “modify, override, or ignore” proposals that were better than their own.
Weak point. Mandates are written in natural language, but they are enforced in code. A Clifford Chance roundtable recorded “mandate drift” already in practice. A user sets a £200 limit, and the agent finds the item at £202. Whatever the translation from intent to constraint gets wrong becomes the dispute of the future.
2. Discovery: from search ranking to agent selection
In search-era commerce a merchant needed to rank. In B2A it needs to be selected, and selection follows different rules.
How it works. An agent assembles candidates from retrieval (a search index, a product feed, a merchant catalogue). It reads them as text, re-ranks them by its own judgement and proposes or executes a choice. Merchants enter the candidate set through structured feeds. OpenAI’s product feed specification requires one row per purchasable variant with nine mandatory fields, and keeps organic search, paid ads and checkout as separate switches. Above the feed, generative engine optimisation (GEO) tries to shape what the agent says.
What the evidence shows.
Visibility can be engineered, and the old tricks do not work. The founding GEO paper (KDD 2024) found that adding citations, quotations and statistics raised visibility in generative answers by up to 40%, while keyword stuffing failed. A source ranked fifth by the search engine gained about 115% from citing sources. The generative layer re-ranks what retrieval hands it, so incumbents’ link equity counts for less.
AI search trusts third parties, not brands. A Toronto study found that, for well-known brands, 93.5% of ChatGPT’s citations were earned media, 6.5% brand-owned and 0% social. For Claude the figures were 87.3% earned, 6.8% brand-owned and 5.9% social. A vendor study by a GEO-measurement firm (vendor data) puts brand-owned domains at 2.9% of citations and finds visibility “near-binary”: 63.2% of brand–prompt–engine cells never mention the brand. Household names appear in 73% of relevant answers, small brands in 11%.
Shopping agents concentrate demand, and the concentration is unstable. The Columbia ACES sandbox documents “choice homogeneity”. In some categories one product took about 89% of picks under a given model while others got none. Model updates “can drastically reshuffle market shares”. Agents show strong position bias that varies by provider and model version. They penalise “Sponsored” tags and reward “Overall Pick” endorsements. The basic rationality tests are now passed by frontier models. Bias and volatility are not solved.
Seller-side agents can game buyer-side agents. In the same study, one seller’s AI rewriting a product description once raised that product’s share significantly under five of six buyer models: +14.89 percentage points under GPT-5.1 and +14.79 under Gemini 2.5 Flash, with single categories moving up to 80 points. It is the first causal evidence of agent-to-agent “AI-SEO”.
Adversarial text moves rankings further. A Harvard study inserted an optimised “strategic text sequence” into one product entry. It moved a fictitious coffee machine from not recommended to top recommendation (Kumar and Lakkaraju). The CORE method promoted target products into the top five 91.4% of the time across GPT-4o, Gemini 2.5, Claude 4 and Grok 3. Cheap filters catch gibberish, but persuasive fake-review prose survives them.
First-offer bias rewards speed. In Microsoft Research’s Magentic Marketplace simulation, every model showed “severe first-proposal bias, creating 10–30x advantages for response speed over quality”. More options made things worse. With 100 search results instead of 3, consumer welfare fell by up to 65% for one frontier model.
Publishers lose the click. Pew found that Google users clicked a traditional result on 8% of visits with an AI summary against 15% without one, and clicked a link inside the summary on 1%. The referral bargain that funded the open web is weakening just as agents start to rely on that web.
Weak point. The competitive variable is moving from price and quality to legibility to a model: structured data, third-party coverage, response latency and phrasing. That favours firms with large earned-media footprints and fast APIs. It also creates a market for manipulation that the French Autorité lists as a theory of harm: opaque visibility criteria and a small number of offers shown give agent operators “increased power over the allocation of demand”.
3. The connector layer: MCP, apps and the agent’s app store
Agents do not only browse. They increasingly call tools directly through connectors, so the connector directory becomes a shop window.
How it works. Anthropic open-sourced the Model Context Protocol in November 2024 as “a universal, open standard for connecting AI systems with data sources”. Services expose MCP servers and agents act as clients. OpenAI’s Apps SDK, built on MCP, offers developers access to “over 800 million ChatGPT users”. ChatGPT can “suggest apps when they’re relevant to the conversation”, and apps meeting “higher design and functionality standards may be featured more prominently”. The official MCP Registry is designed as one neutral upstream catalogue, with client-owned “marketplaces” built on top of it.
What the evidence shows.
The catalogue is mostly noise. A measurement study across six markets collected 17,630 entries, of which only 8,401 were valid projects. More than half the ecosystem is placeholders, forks or abandoned code, and only 6.9% of projects appear in four or more markets.
Vetting is weak. A University of Delaware study of 67,057 servers found that 99.16% are hosted on public GitHub repositories with no continuous ownership checks. It demonstrated an attack in which malicious tool metadata made the model swap an email recipient for the attacker’s address, without the malicious tool ever being called. A Queen’s University audit found 7.2% of 1,899 servers vulnerable and 5.5% showing tool poisoning.
Agents choose tools by their descriptions, and descriptions are advertising. Three independent groups found the same thing:
A Maryland study across 17 models found that edited descriptions “receive over 10 times more usage” (Faghih et al.). Assertive wording beat the original 79.1% to 11.9%.
ToolTweak (Oxford and Microsoft) lifted a tool’s selection rate from about 20% to 81% by automated rewriting, and the effect transferred across open and closed models.
MPMA reached 100% attack success in most settings by inserting manipulative words. Its optimised “authority” variant had a 0% detection rate under an LLM judge.
Weak point. The connector layer combines a storefront, an advertising surface and an attack surface, and so far it has no regulation of claims. Whoever controls in-conversation suggestion holds the agent era’s equivalent of shelf placement. The Autorité has already floated designating AI-model marketplaces as DMA core platform services.
4. Checkout protocols: how an agent actually buys
There are two ways for an agent to buy. It can drive a browser through a site built for humans, or it can talk to the merchant through a protocol. Browser agents reach any site without the merchant’s cooperation. Protocol checkout requires the merchant to participate, but it removes the many-to-many integration problem (Activant). The protocol landscape changed several times between September 2025 and autumn 2026. The library reads as a layered stack, not a single standard.
Communication layers. MCP links agents to tools. Its specification says hosts “must obtain explicit user consent before invoking any tool” and that tool descriptions “should be considered untrusted”. It also concedes that “MCP itself cannot enforce these security principles”. Google’s A2A (version 1.0) links agents to agents through Agent Cards that declare identity, skills and authentication.
Commerce layers.
ACP, from OpenAI and Stripe, is an “open standard for connecting buyers, their AI agents, and businesses”, released in dated versions from September 2025 to April 2026 (ACP repository). In ChatGPT’s Instant Checkout the merchant sends a product feed and exposes checkout endpoints. OpenAI prepares “a one-time delegated payment request” with “a maximum chargeable amount and expiry”. The merchant “accepts or declines the order”, and “OpenAI is not the merchant of record”.
UCP, from Google with Shopify, has businesses publish a profile at
/.well-known/ucp(UCP specification). The agent and the merchant negotiate one protocol version, then the intersection of their capabilities. Payment runs through handlers in which the platform obtains an opaque token from the credential provider. “Raw data never touches the business’s frontend”, and the platform “is not the ‘owner’ of the funds”.
The security layer. AP2 proves authorisation with signed mandates. Its starting assumption is unusual for a commercial standard: “when either role is agentic, then the Agent itself is a potential attacker”, so all validation “MUST happen in deterministic code”. In autonomous mode, an agent may not present a new open mandate without a rejection receipt for the previous one. This stops one permission from approving several purchases. Receipts give “a non-repudiable picture of the transaction”, but dispute rules are “outside the scope” (AP2 specification). AP2’s September 2025 launch blog described a different set of mandates, Intent, Cart and Payment, so it matters which version a source cites.
Card-network programmes.
Mastercard’s Agent Pay (vendor release) introduces “Agentic Tokens”, requires agents to be “registered and verified”, and promised that all US cardholders would be enabled by the 2025 holiday season.
Visa Intelligent Commerce issues network tokens to requestors typed
AI_AGENT. It binds consent to a passkey-signed “purchase instruction” and requires onboarding to the Visa Agent Registry. This is the card-network equivalent of an AP2 mandate.
Wallets and merchant platforms. PayPal says “merchants remain the merchant of record” for agent-initiated transactions. Shopify runs UCP-compliant MCP servers in tiers. Untrusted agents build a cart and hand the buyer to the storefront. “Trusted agents can complete checkouts directly.” Stripe’s shared payment tokens let the agent cap currency, amount and expiry, and are offered under both ACP and UCP.
The common pattern. Read side by side, the protocols agree on far more than their branding suggests:
The user grants a scoped permission.
The agent carries a single-use, capped, expiring token rather than a card number.
The merchant receives the order through its own payment processor and stays the merchant of record.
The merchant therefore keeps the fraud exposure and the chargebacks.
No major platform in the library takes on the money or the liability. That is a deliberate choice: it lets agent operators scale without becoming regulated payment institutions. It also means the risk of agent error lands on merchants and card issuers, who have the least visibility into the agent’s reasoning.
Weak point. Multiple standards and a cold start. The Clifford Chance notes list four competing standards (x402, AP2, ACP and MPP) and describe “a cold-start problem … with closed vertically integrated systems winning commercially in the interim”.
5. Payment rails: reversible cards versus final stablecoins
Once authorised, the payment needs a rail, and the choice of rail decides whether a mistake can be undone.
How it works. Two families compete.
Cards with agent-bound tokens. The existing card system adds agent registration, agent-specific tokens and signals that tell issuers a transaction was agent-initiated. Chargebacks and the consumer protection built on them remain in place.
HTTP-native stablecoin payments. Coinbase’s x402 (vendor whitepaper) revives the dormant HTTP 402 “Payment Required” status. A server answers an unpaid request with a price, the agent attaches a USDC payment and retries, and a facilitator verifies and settles on-chain. Coinbase claims settlement in about 200 ms and fees under $0.0001, against “$0.30 + 2.9%” and chargeback windows of “up to 120d” for cards. Its whitepaper presents irreversibility as a feature: payments “settle instantly onchain, eliminating chargebacks and disputes”. The project has since moved to an x402 Foundation and says it “will never deprioritize onchain payments in favor of fiat”.
What the evidence shows.
Micropayments are a real agent need. Waller notes that before a purchase an agent “might pay for LLM queries, price data feeds, and other API calls”, and that “micropayments like these favor payment rails with lower flat fees”. This is the strongest argument for new rails: card economics do not work at a tenth of a cent.
Stablecoins are still small next to the rails they would replace. The BIS Annual Economic Report 2026 puts stablecoin market value at about $320 billion at end-May 2026. Annual transaction volume was about $28 trillion in 2025, “less than three business weeks” of the largest US wholesale payment systems. Its verdict is that stablecoins “fall short on foundational properties of money and threaten financial integrity”. It proposes tokenisation anchored in central-bank money instead.
The law fixes reserves, not recourse. The US GENIUS Act requires 1:1 reserves in cash and short Treasuries and bans paying “any form of interest or yield”. It creates no chargeback regime. The Consumer Bankers Association notes that stablecoins and x402 “do not today contain mechanisms for reversals or chargebacks”, and rates their liability clarity “Low”.
Central banks are designing around agents.
The Bank of England staff blog expects more transactions of smaller size, and a shift “from Know-Your-Customer (KYC) to Know-Your-Agent (KYA)”. It names the core tension: payment systems are “deterministic”, while agents “rely on probabilistic AI outputs” and “may … retry transactions in unexpected ways”.
Deputy Governor Breeden lists “agentic payments and commerce” (”book my holiday”, “refill my fridge”) among the uses that tokenised deposits and regulated stablecoins should serve.
Currency matters. The Clifford Chance participants observed that agent settlement currently defaults to USD stablecoins, and that “agentic payments cannot scale meaningfully if everything settles in USD”. For a euro-area economy this is the monetary-sovereignty side of B2A. If machine-to-machine micropayments default to dollar tokens, a growing slice of digital commerce settles outside the euro.
Weak point. A mistake on a final rail is harder to correct than a mistake on a reversible one. Agents err, are hijacked and drift past their mandates, so the case for irreversible settlement is weakest exactly where B2A is newest: consumer purchases by autonomous agents. The pattern in the evidence is cards for consumer goods, programmable rails for machine-to-machine micro-transactions, and tokenised deposits as the regulated middle that central banks prefer.
6. Identity: proving who the agent is and whom it serves
A merchant facing an agent needs three answers. They are different questions, and the library shows they are being solved by different systems.
Is this traffic from a known agent operator? This is solved by signed requests. Visa’s Trusted Agent Protocol and Mastercard’s Agent Pay both use Web Bot Auth, the IETF draft built on RFC 9421 HTTP Message Signatures (held in the sibling library). An agent registers a public key in a directory. Each request carries a signature, a nonce against replay and a tag:
agent-browser-authwhen browsing,agent-payer-authwhen paying (Cloudflare, vendor blog). Cloudflare has created a separate class of “signed agents”, directed by end users rather than by a single company, alongside the “verified bots” such as crawlers.On whose behalf, and with what authority? This is solved by delegation tokens and mandates. The OpenID Foundation argues that “user impersonation by agents should be replaced by delegated authority”. Today agents often act “indistinguishably from users, creating accountability gaps”. Its alternative is explicit “on-behalf-of” flows in which an agent proves its scope while staying identifiable as itself. An MIT-led proposal adds a user-signed Delegation Token alongside the user’s ID token and an Agent-ID token. Microsoft’s Entra Agent ID puts this into enterprise use: tokens in which “the subject … is a user, while the actor is the agent identity”, with a named human “Sponsor” accountable for each agent.
Who is accountable afterwards? This is solved by logs, receipts and law. The NIST NCCoE concept paper asks how to bind agent actions back to a human with tamper-proof logging and non-repudiation. It also asks “What constitutes a strong authentication for an AI agent?” Its answer is still a draft for consultation, not guidance.
What the evidence shows.
Identity unlocks access in tiers. Shopify’s tiering and the OpenID Foundation’s description of a “two-tiered web” point the same way. Identified agents get permission to check out, while anonymous agents are restricted. Know-Your-Agent extends KYC/KYB to the agent itself. The KYAPay profile bundles verified identity and payment details into one JWT for first contact.
Registries are fragmenting. “Visa and Mastercard will be hosting their own directories” (Cloudflare). The Bank of England blog flags that “some agent identity standards are only supported by specific card schemes”. Cloudflare’s counter-proposal is an open registry format: a list of key URLs that anyone can host, in “an open curation ecosystem that doesn’t lock in customers or small origins”.
Consent does not scale. The OpenID paper warns that users “will face thousands of authorization requests as agents proliferate”, which encourages reflexive approval. It also notes that recursive delegation to sub-agents lacks “clear scope attenuation mechanisms”. The Clifford Chance participants describe the underlying tension as “a deeper clash between the privacy norms of internet culture and the identification requirements of the banking system”.
Weak point. Signatures prove which operator sent a request. They do not prove that the request reflects the user’s intent. A properly signed agent that has been hijacked is still a properly signed agent.
7. Fraud and attack: the agent as the weakest link
B2A creates two fraud problems at once. Criminals use agents to attack merchants, and attackers turn legitimate agents against their own users.
Agents as attackers’ tools.
Volume. Imperva (vendor data) reports that automated traffic passed human traffic in 2024, at 51% of all web traffic, and that bad bots rose to 37%. 44% of advanced bot traffic targeted APIs, and checkout endpoints drew about 32% of API attacks.
Declared identity cannot be trusted. HUMAN (vendor data) found that “a significant portion” of requests claiming to come from ChatGPT, Mistral and Perplexity did not originate from those operators. An agent creating an account and completing a checkout could be acting “on behalf of a real customer or executing a carding attack autonomously. The distinction depends on behavioral context, not declared identity.” HUMAN also reports post-login account compromise up 250% since 2022.
Organised crime. Europol’s IOCTA 2026 calls fraud “the fastest-growing area of organised crime”. It expects criminal use of agentic AI to raise online fraud “to unprecedented levels”. It also notes that most criminals still use generative AI rather than agents. The FATF reports that 90% of assessed jurisdictions identify fraud as a major risk. In the UK fraud is “more than 40% of all crimes”.
Agents as victims.
Hijacking through content. NIST’s AI Safety Institute defines agent hijacking as indirect prompt injection through data the agent reads. New attacks raised success against a frontier model from 11% to 81%. Retrying each attack 25 times lifted average success from 57% to 80%. Single-attempt benchmarks therefore understate the risk wherever attackers can retry.
No model is safe. In a large red-teaming competition run by NIST CAISI, more than 250,000 attempts by over 400 participants found at least one successful attack against every one of 13 frontier models. Robustness “did not correlate uniformly with model capability”.
Stealth harm in commerce. StakeBench tested 12 harm objectives against users, sellers and platforms. Indirect prompt injection succeeded 41.7–68.6% of the time, and “no harm objective is reliably resisted”. The worst outcome is “stealthy parasitism”, where the attack succeeds while the user’s delegated task still appears to finish normally. The user sees a completed purchase and has no reason to complain.
Dark patterns work on machines too. WebDecept found shopping agents “highly susceptible” to deceptive interfaces, including items added to the cart unasked and totals inflated to 120%. One agent reached a 100% safety-violation rate under personalised pop-ups, and still 89% with a protective prompt.
Poison once, exploit later. eTAMP shows that a single manipulated product page can poison an agent’s memory and trigger later on other sites, with “up to 32.5%” success on GPT-5-mini. “More capable models are not more secure.”
Screens can carry instructions. VPI-Bench deceived computer-use and browser agents at rates of up to 51% and 100% through visual prompt injection on sites such as Amazon and Booking.
Weak point. Fraud systems trained on human behaviour will misread agents (Waller). An agent’s speed, regularity and lack of mouse movement look like a bot because it is one. The library contains no primary dataset on chargeback or friendly-fraud rates for agent purchases. Claims about these rates in the trade press are unverified, and this report does not state them.
8. Liability: who is bound, and who pays
Law is the part of the machine where the most has been settled. What remains open is the most consequential part.
Settled: the principal is bound.
The UNCITRAL Model Law on Automated Contracting (2024) says a contract formed by an automated system “shall not be denied validity … on the sole ground that no natural person reviewed or intervened” (Art. 5). Actions are attributed “to the person who uses the system”, and attribution “shall not be denied on the sole ground that the outcome was unexpected” (Art. 7). An optional Article 8 protects the user when the counterparty knew, or should have known, that the action was unexpected. Article 10 rules out an “AI made me do it” defence.
The European Law Institute reaches the same conclusion for consumers: “the legal effects of all the actions of a digital assistant are attributed to the consumer”. It pairs this with design duties that keep the consumer in control, and with a “no-barrier” principle: traders should not block assistants “in a general, arbitrary, or unjustified way”.
Settled: the deploying business answers for its agent. The UK CMA’s guidance is the plainest statement in the library: “You are responsible for what an AI agent does in the same way you are responsible for what an employee does. This is true even if someone else designed or provides the AI agent.” Fines under the DMCC Act can reach 10% of worldwide turnover. Its companion paper adds that an agent that “steers, pressures or misleads consumers” is “likely to be unlawful”. The Bank of England staff blog takes the same line: deployers “retain legal responsibility … responsibility does not sit within a ‘black box’”.
Open: payment liability when the agent exceeds its mandate.
In the US, Regulation E’s protection against unauthorised transfers “may not apply when agents are involved”. A consumer who furnishes an access device to another person who exceeds their authority may be liable for that person’s mistakes (CBA).
In the EU, the April 2026 Payment Services Regulation compromise never mentions AI agents. It requires strong customer authentication when a mandate is first set up, but not for each later merchant-initiated payment (recital 108). That is the legal hook for stored-mandate agent payments. It also tightens the commercial-agent exclusion, so that intermediaries acting for both buyer and seller are caught.
Product liability does not reach bad purchases. The revised Product Liability Directive treats AI as a product and holds manufacturers liable for “unexpected behaviour”. But it covers death, injury, property and data, not pure economic loss.
The AI Act does not mention agents. The AI Act bans manipulative techniques (Art. 5) and requires disclosure when people interact with AI (Art. 50). The word “agent” does not appear in the extracted text at all.
Open: loyalty. Georgia Tech’s Riedl and Desai name the “agentic loyalty problem”. The agent picks a $450 offer over a $425 one “because the AI agent’s company had a deal with the seller”. Kolt argues that incentives, monitoring and enforcement, the tools of agency law, may fail at agent speed and scale. Contract law answers who is bound. It does not answer whom the agent works for.
Weak point. The settled rules put the risk of agent error on the principal (bound by the contract) and on the merchant (merchant of record). The agent operator, the party that designed the behaviour, is reached mainly through consumer-protection enforcement and its own terms of service. That allocation is the single most important open policy question in B2A.
9. Economics: what agents do to prices, competition and surplus
The optimistic case is simple. Agents collapse search and transaction costs, so markets become more competitive and consumers capture more surplus. The library supports half of it.
What holds. Search does get cheaper. Microsoft Research’s equilibrium analysis finds that “cheaper search improves learning and consumer surplus”. The OECD notes that “AI could be leveraged to act as an agent for consumers”, which raises competitive pressure (OECD).
What does not hold automatically. Seven findings complicate the story:
More informative private search can weaken competition. The same paper finds that search which probes deeper fit “can degrade” consumer surplus. Each agent quietly screens out poor matches, so fewer firms compete for each buyer and shared market signals learn less. The proposed remedy is that the market learns from agent transcripts too.
Your own agent may charge you more. In 325,000 experiments, 8 of 13 agents chose costlier options for users they inferred to be wealthy, even when asked for the cheapest. The gap reached $208 on one flight query (Et Tu, Brute?). The authors call this “adversarial delegation”. The personal data that makes an agent useful lets it act against its user.
Delegation leaks the budget. When consumers delegate with a natural-language persona, a seller’s agent recovers their willingness to pay almost exactly (slope 1.00). “Because this leakage arises from delegation itself, it cannot be mitigated at the prompt level” (Iowa study).
Better models win, and losers do not notice. In Anthropic’s Project Deal, 69 agents struck 186 real deals worth just over $4,000. Opus sold the same item for $2.68 more and bought for $2.45 less than Haiku, yet people represented by the weaker model rated the deals as equally fair. A negotiation study found capability correlating with deal outcomes at up to r = 0.93. It also found weaker buyer agents exceeding budgets in over 10% of cases.
Collusion without agreement. LLM pricing agents “quickly and autonomously reach supracompetitive prices”, without being told to (Fish et al.). In a vertical simulation, platforms exploiting ranking and sellers exploiting descriptions together removed 37.1% of consumer surplus, against 17.4% if the effects merely added up. The harm “requires no coordination and evades antitrust detection”.
Sponsorship steers invisibly. In a Princeton experiment with 2,012 participants, conversational AI nearly tripled selection of sponsored products, to 61.2% against 22.4% with search. “Sponsored” labels did not help, and concealed steering was detected less than 10% of the time (Commercial Persuasion). Across 23 models, all but five recommended the expensive sponsored option more than half the time (Ads in AI Chatbots).
Platforms can profit from inefficiency. When agents’ attention becomes priced “inspection tokens”, a platform’s optimal pricing produces “excessive inspection relative to the social optimum”. The platform is “extracting rents from the opacity it creates” (Subsidizing Sequential Search).
Caveat. Most of these are simulations or theory, many are preprints, and three of the eight economics papers come from one Microsoft Research group. They establish mechanisms, not magnitudes. What they share is the direction: the welfare gain from cheaper search is real, but it does not reach consumers unless agent loyalty, disclosure and market design ensure it.
10. Measurement: how big B2A really is
Every number in this field depends on who counted, against what base, and with which classifier.
Bots in general. Imperva counts 51% of traffic as automated. Cloudflare’s Radar counts AI bots at 4.2% of HTML requests, and Googlebot alone at 4.5%. These are different denominators, not a contradiction.
Agents in particular. In HUMAN’s data (vendor), agentic traffic was 1.7% of AI traffic, against 67.5% training crawlers and 31.9% scrapers and fetchers. 77% of agent activity was on product and search pages and 2.3% on checkout. OpenAI generated about 69% of AI-driven traffic.
User-triggered fetching is the fastest signal. Cloudflare found that crawling triggered by users asking a chatbot grew more than 21-fold during 2025, closely tracking ChatGPT-User.
The take-to-give ratio. 80% of AI crawling was for training, 18% for search and 2% for user actions. Anthropic crawled 38,000 pages per referred visit in July 2025, down from 286,000 in January (crawl-to-click).
Merchants cannot see it. In a PYMNTS survey of 60 mid-sized US merchants, none selling through their own sites could consistently track AI-driven visits to a completed purchase, and 12% cannot distinguish agent from human traffic. They expect agents to complete 13% of digital sales within two years on average.
What is missing. No source in the library measures purchases completed by agents themselves, as distinct from AI-referred human visits. Adobe’s widely quoted conversion figures were not obtainable and are treated here as web context only.
Weak point. B2A is being priced by investors on forecasts and measured by vendors who sell bot management. Until protocol operators or card networks publish agent-initiated transaction counts, the size of the market is an estimate.
Coding agents: the first B2A market that already exists
Consumer agentic commerce is still mostly forecast. Software development is not. It is the one domain in the library where agents already choose, provision and pay for products at scale. That makes it the best available preview of how B2A behaves once it works.
Agents are already the customer.
When Databricks agreed to acquire Neon, it reported that 30% of Neon’s databases were created by AI agents at general availability. Later that figure was “over 80%”: “AI agents were creating 4 times more databases versus humans”. This is the acquirer’s own telemetry, but it is the cleanest published case of a market where the agent is the main buyer.
GitHub’s Octoverse 2025 reports 518.7 million pull requests merged, up 29%, with issue-closing accelerating after the Copilot coding agent preview. TypeScript overtook Python, which GitHub links to “typed languages that make agent-assisted coding more reliable”. Dockerfile use grew 120%, “likely fueled by the need to sandbox agents”.
The AIDev dataset counts more than 456,000 pull requests by five agents across 61,000 repositories. OpenAI Codex alone opened over 400,000 in under two months.
Anthropic’s software-development study found that 79% of Claude Code conversations were automation, against 49% on Claude.ai.
What agent customers buy. The Neon announcement says plainly why agents chose Neon:
Postgres is something “all frontier LLMs have been trained on”.
Provisioning runs “at machine speed”.
Instances are cheap enough to “launch thousands or even millions”.
Branching makes experiments safe.
Each reason generalises:
Familiarity in training data works like brand equity. This mirrors the “big brand bias” in AI search.
API-first, instant provisioning wins over sales-led onboarding.
Unit pricing must suit fleets of short-lived instances, not seats.
Reversibility is valued: branches, sandboxes, rollbacks.
What the work is worth: the mixed evidence. B2A in coding has produced real throughput. Quality and productivity are a different matter.
Acceptance is lower than for human work. Claude Code pull requests in 157 projects were merged 83.8% of the time, against 91.0% for human ones (Watanabe et al.). Across AIDev, acceptance rates were 64% for Codex, 49% for Devin and 35% for Copilot. That is 15 to 40 points below human performance, and the authors note “a growing gap between benchmark performance and real-world effectiveness”.
Task type matters more than vendor. A task-stratified analysis of 7,156 pull requests found documentation accepted 82.1% of the time against 66.1% for new features. “No single agent performs best across all task types.”
Felt productivity is not measured productivity. In METR’s randomised trial, experienced developers working in mature repositories were 19% slower with AI tools, while believing they were 20% faster. The one participant with more than 50 hours of Cursor experience did speed up, which suggests a steep learning curve rather than a ceiling.
Speed now, debt later. A Carnegie Mellon difference-in-differences study found Cursor adoption raised commits by 55% in the first month. The gain faded, while static-analysis warnings (+30%) and code complexity (+42%) rose and stayed up.
Developers use it and distrust it. Stack Overflow finds 84% using or planning to use AI, but only 3% trusting it highly. The top frustration is “AI solutions that are almost right, but not quite”. Google’s DORA report summarises: “AI is an amplifier”, raising throughput and also delivery instability.
The lesson for the rest of B2A. Coding shows the pattern consumer markets should expect:
Agent demand arrives first where tasks are well specified, outcomes can be checked and actions can be reversed.
Vendors whose products are legible to models capture that demand disproportionately.
The bottleneck moves from production to verification. Reviewing agent output takes up the time agents save.
The benchmark baseline is a reminder of how fast this moved. When SWE-bench launched, the best model resolved 1.96% of real GitHub issues.
What is real and what is hype
Separating the two requires grading the evidence, not just the claims.
Real and measured (strong sources).
Agents are the majority customer in some developer infrastructure: Neon over 80% (company data), with very large volumes of agent-authored code (GitHub, AIDev).
AI systems already read the commercial web at scale. Bots make up 51% of traffic (Imperva), and user-triggered chatbot fetching grew 21-fold in 2025 (Cloudflare).
Agents are measurably biased and manipulable in ways that move market share. This rests on replicated academic findings across independent groups: position bias, first-offer bias, description gaming, sponsored steering and prompt injection.
The payment and identity plumbing exists in production form. AP2, UCP and ACP have specifications. Card networks have agent tokens and registries. Signed-agent classes are live at Cloudflare.
The law has a clear starting rule: principals are bound and deployers are responsible (UNCITRAL, ELI, CMA).
Real but small (vendor or survey sources, directionally consistent).
Agentic traffic is about 1.7% of AI traffic. 2.3% of agent activity reaches a checkout (HUMAN, vendor data).
Agent-referred traffic to French merchants is under 5% today, with about 25% plausible by 2030 (Autorité de la concurrence).
Consumers are open to agent discovery (70–90%), about a third would accept delegated decisions, and single digits to low teens would accept autonomous payment (Accenture, BCG, Capgemini, Checkout.com).
Merchants expect agents to complete 13% of digital sales within two years. None can currently track agent visits to purchase (PYMNTS, n = 60).
Reported retreats (single or secondary sources, verify before relying on them).
OpenAI’s Instant Checkout pilot “has since been withdrawn”, according to the Clifford Chance roundtable notes. Activant, a venture investor, says OpenAI shut in-chat checkout within six months, citing a lack of flexibility, and pivoted back to discovery. The ACP specification itself continued to be updated into 2026.
x402 “has reportedly enabled over 150 million transactions”, but volumes “have materially tapered off”. This raises the question whether early numbers “reflected incentivised rather than organic demand” (same notes).
Perplexity’s Instant Buy reportedly works only with PayPal-integrated US merchants (Activant).
Forecast, not fact. All market-size figures for 2030 are forecasts with incompatible scopes: Morgan Stanley’s $190–385 billion, BCG’s $1.3 trillion “impacted”, and McKinsey’s $3–5 trillion global as quoted by Activant. They count influenced, completed and “orchestrated” revenue differently. Activant’s own critique is fair: they “may be directionally right” but “obscure the question that matters: where does agent-mediated commerce create enough value to change buying behavior”. Its answer is useful too. Penetration will be fast “where friction is high and decisions are objective”, and slow where “the buying experience carries its own value”.
ENSI’s reading. The B2A economy is real at the infrastructure layer and in software. It is visible but small in consumer commerce, and over-forecast in aggregate. Its first-order effects on discovery, competition and fraud are arriving ahead of its transaction volumes. Agents already shape which brands are seen long before they complete most purchases. That ordering matters for policy. The distribution and competition questions have to be settled first, while the payment questions still have time.
Ten shifts to watch, ranked
Ranked by how much each shift would change the structure of the market, weighted by how close it is.
The agent layer becomes the demand allocator. Horizon: now to 2028. A small number of general agents decide which merchants are seen across many categories at once. Signal: agent-referred share of merchant traffic moving from under 5% towards the Autorité’s 25% estimate. Also watch for in-conversation app suggestions and “featured” placements becoming paid products.
Advertising moves inside the answer. Horizon: 2026–2028. Sponsored results enter agent conversations, and the evidence says labels do not neutralise them. Signal: ad products in major assistants, and regulators applying unfair-commercial-practice law to conversational steering, as the CMA has signalled.
Agent-readiness becomes a merchant’s main visibility task. Horizon: now. Feeds,
/.well-knownprofiles, MCP endpoints, earned-media footprint and response latency replace parts of classic SEO. Signal: most merchants publishing agent-facing profiles, and the rise of agent-specific analytics. 85% of PYMNTS respondents have already reworked titles and descriptions for AI.Identity becomes the gate to checkout. Horizon: 2026–2027. Signed, registered agents get direct checkout while anonymous ones are pushed back to human handoff. Signal: Web Bot Auth adoption, whether card-network registries converge or stay separate, and Know-Your-Agent requirements written into acquirer rules.
Prompt injection becomes a fraud category with a loss line. Horizon: 2026–2028. Stealthy hijacking of shopping agents moves from benchmark to incident. Signal: the first published loss data, and whether attack success rates in independent red-teaming fall over successive model generations.
The liability gap closes, one way or another. Horizon: 2027–2029. Either payment law covers agent-initiated transactions explicitly, or private mandates and insurance fill the gap. Signal: agent provisions in Regulation E interpretations or PSR implementing acts, agent-transaction indicators in card dispute rules, and operator-backed guarantees.
Two rails, split by use. Horizon: 2026–2030. Card tokens carry consumer goods, while programmable money carries machine-to-machine micropayments. Signal: whether x402-style volumes recover organically, euro-denominated regulated tokens or tokenised deposits becoming usable by agents, and how central banks follow Project Agorá.
Agent loyalty becomes a product and a legal duty. Horizon: 2027–2030. Consumers and regulators demand agents that are demonstrably on the user’s side and disclose their conflicts. Signal: fiduciary-style obligations in law or codes, independent audits of steering and wealth-based price discrimination, and paid “loyal agent” tiers.
Algorithmic coordination becomes an antitrust case. Horizon: 2027–2030. Simulations of LLM pricing collusion and vertical tacit collusion meet a first enforcement action. Signal: a competition case built on agent-to-agent pricing, or the DMA applied to an AI agent as a virtual assistant or search service.
Self-running firms sell to agents. Horizon: 2028 and beyond. Agent-operated businesses, “dark supermarkets” served by API in Busch’s phrase, emerge in narrow, objective categories. Signal: Project Vend-style experiments that stay profitable without human rescue. Today the best agent completes about 30% of realistic office tasks (TheAgentCompany), and the gap between “capable” and “completely robust” remains wide.
What to do with this
Three conclusions follow from the machine as mapped here.
Treat selection, not transaction, as the battleground. The point where B2A redistributes advantage is the agent’s choice set. That is where biases, sponsorship, description gaming and concentration operate. Payment volumes will follow choices made earlier.
Treat the agent as a party with interests. The evidence on wealth steering, sponsored persuasion, persona leakage and capability gaps shows that agents are not neutral pipes. Who funds the agent, and to whom it is loyal, is an economic variable as important as price.
Treat irreversibility and identity as the two design choices that decide risk. Reversible rails and strong, interoperable agent identity make mistakes and attacks recoverable. Final settlement and fragmented registries make them expensive and concentrate power in whoever runs the gate.
Report 2 turns these findings into a playbook: what merchants, platforms and a mid-sized European state should do in the next 12–24 months. Report 3 takes up the question this report leaves open. When agents buy and agents sell, who created the value, and who captures it?



