Agent economy sources and evidence

Adobe reports 62% year-on-year growth for retail in its August report.

Visit-share growth is not an absolute spending share or delegated execution.

Travel AI referrals grew 119% year on year. The AI/non-AI conversion gap narrowed to 1%.

Visits and conversion do not identify who executes a booking.

Adobe reports 31% year-on-year growth in financial-services AI visit share.

A research visit is not an agent-executed financial transaction.

NIQ reports about 34% use AI for product research, while about 8% have allowed it to complete a purchase.

Consumer incidence is not the share or value of purchases. Geography, fieldwork and autonomy need fuller methodology before AEI use.

Mastercard introduced payments for high-frequency machine transactions, supported by more than 30 industry participants.

Digital-services and logistics examples are proposed use cases. This is not measured deployment, transaction value or AEI.

Mastercard describes work with partners to enable sourcing, ordering and payment for businesses in Australia and New Zealand.

The article gives no country-specific transaction volume or value. Planned enablement is not measured economic delegation.

Claude agents represented 69 employees and negotiated 186 real deals without further human sign-off during negotiation.

A self-selected, subsidised internal marketplace cannot estimate the national T5 share. Physical goods were subsequently exchanged by people.

Visa describes permissioned credentials, agent identity and payment controls as foundations for agentic commerce.

Network-wide transaction volumes are not agentic transaction volumes. No AEI numerator is disclosed.

Visa launched a programme for issuers to test and validate agent-initiated payments, starting in Europe including the UK.

Readiness testing is not transaction adoption. No country-level autonomous spending value is disclosed.

Santander tested an AI-initiated payment through live infrastructure with Mastercard Agent Pay.

The publisher explicitly says this is not a commercial rollout. No national allocation or transaction-value share is available.

Ant Group reports that Alipay AI Pay surpassed 100 million users during Chinese New Year.

Provider-reported users are not transaction value. The release does not establish the autonomy of every purchase or a market-wide AEI denominator.

Ant Group reports more than 120 million AI Pay transactions during 5–11 February 2026.

Payment count is not transaction value. Voice-command payment does not establish autonomous choice for every purchase. This overlaps the same release’s user figure and is not independent evidence.

Salesforce attributes $262 billion, about 20% of global online holiday sales, to AI and agent influence.

Includes recommendations and customer engagement. Influence is broader than AEI T3 and is not proof of delegated execution; do not insert 20% into the formula.

Adobe measured 693.4% growth in traffic arriving from generative AI tools during the holiday period.

This is traffic growth from a modest base, not 693.4% adoption. Different periods and definitions make comparison with the July visit-share metric unsafe.

PayPal enabled US Perplexity users to check out with merchants inside the shopping conversation.

The published flow includes user clicks. In-chat checkout does not automatically satisfy T4 autonomous choice and execution; no value share is disclosed.

EMVCo opened a draft framework for managing persistent consumer intent in card-based agentic payments, including scenarios where a consumer is absent when payment is executed.

This is a draft technical framework. It reports no live transaction count or value and cannot populate any AEI tier.

NIQ and Similarweb plan to connect AI discovery, traffic and conversion signals with measured retail sales, with an initial version expected in Q4 2026.

The release is a product forecast, not a dataset. It does not separate autonomous execution from human-confirmed purchases or provide transaction-value shares.

An AI agent monitored product availability and completed a PLN 19.99 purchase after the user had previously set conditions and authorised payment; no user action was required at execution.

This is one pilot transaction, not a representative series or market-value share. Prior user authorisation means the event is relevant to delegated execution, but it supplies no denominator for Europe or Poland.

Mastercard announced Agent Connect and expanded merchant tools covering product discovery, recommendations and a trusted path to purchase across participating agents.

The release describes capabilities and partner intentions. It does not report production transaction value, count or autonomy mix.

Ant International says ten Alipay+ wallets serving 1.5 billion accounts and seven acquirers are participating in Phase I of its Agentic Mobile Protocol rollout.

Partner reach and eligible accounts are not active users or transactions. The release gives no completed agentic-payment value or mutually exclusive autonomy tiers.

Meta launched Muse in the United States with browser-based shopping, negotiation and checkout through Stripe Link one-time-use cards; Meta says the agent returns for approval before making a purchase.

The launch reports capability, not transaction count or value. Because a person approves before purchase, the flow cannot be treated as autonomous T4 execution without transaction-level evidence on the agent's decision authority.

Visa reports that 23% of surveyed US consumers trust generative AI to handle payment transactions on their behalf; the relevant question base was about 1,028 to 1,034 respondents.

Stated trust is not observed use, purchase incidence or transaction value. The result cannot be mapped to T4 or T5 and the publication date is later than the May fieldwork.

ACI Worldwide reports that 7% of more than 3,000 UK and US fashion and sportswear shoppers would let an AI assistant buy without per-purchase approval, split between set limits and specified retailers.

This is stated willingness in one retail category, not observed delegated purchasing or transaction value. The source page does not publish country-level splits or fieldwork dates.

Amazon says shoppers can instruct Alexa for Shopping to buy a specified item automatically when it reaches a target price, using the default payment method.

The article reports availability, not completed purchases, transaction value or geographic coverage. A feature launch cannot establish market share or a comparable T4 series.

Mastercard, Trip.com and Network International are showcasing a flow intended to let users search, book and complete purchases through TripGenie, with the solution expected in early 2027.

This is a collaboration announcement and future availability forecast. It supplies no live booking count, value, autonomy breakdown or country allocation.

Salesforce and Google announced a native product-feed integration and UCP-powered checkout, with the UCP integration scheduled for general availability in October 2026.

The release is a product and availability forecast, not evidence of completed agentic purchases or transaction-value share.

Mastercard, Flybits and Rogers Bank report Canada's first successful deployment of an agentic commerce transaction. A Flybits assistant identified an eligible product within a Rogers Red cardholder's specified budget and completed the purchase through Mastercard Agent Pay.

The release describes one transaction and does not publish its value, merchant category, population denominator or recurring volume. Consumer-control wording and validated instructions do not establish an economy-wide autonomy share. The accompanying 2,000-adult survey measures openness, not completed purchases.

NIQ's monthly tracker reports that 51% of US consumers used at least one AI-powered tool to support shopping in the past month, including 20% using AI product recommendations and 16% using AI personal shopping assistants.

The release does not state the latest-wave sample size or fieldwork dates. Self-reported support across discovery, comparison and decision-making is not autonomous execution, purchase incidence or transaction value, so it cannot populate an AEI tier.

Anthropic ran a controlled barter market with 201 employees across six offices. Claude-powered agents inferred book preferences, negotiated with other agents and executed bilateral or multi-party swaps when every agent accepted.

This was an employee experiment using books already contributed by participants, not representative commercial activity. No money changed hands, there is no transaction-value denominator, and the sample is not representative. The follow-up willingness to delegate about 30% of book budgets is stated intent, not observed spending.