AI can help you pick products, but do you dare to let it spend your money?

Author: insights4vc

Translated by TechFlow

TechFlow Brief: AI is leading the way as an entry point for commercial interactions, but it still has some distance before becoming a large-scale autonomous buyer. Consumers are accustomed to using AI to research products, compare alternatives, and click through to merchant websites to place orders. The evidence that they truly hand over purchasing decision-making rights to agents remains much weaker.

However, recent data from the Chinese market shows that AI embedded within mature shopping applications can already significantly influence transactions. This distinction shifts the investment logic: opportunity depends on which transaction costs software coordination can reduce, and who controls distribution, authorization, and fulfillment.

Stablecoins enter this discussion because agents require payment mechanisms. However, product discovery, purchase authorization, and settlement infrastructure each operate with independent economic equations. This article breaks down these boundaries step by step using retail data, benchmark tests, and cost projections.

Before AI becomes a significant autonomous buyer, it is first emerging as a meaningful interface for commercial interaction. Consumers use it to research products, compare alternatives, and visit merchant websites when ready to buy. By comparison, the evidence that consumers routinely delegate purchasing decisions to AI remains weak.

Nevertheless, recent results from China indicate that long before independent agents become the primary transaction channel, AI embedded within mature shopping apps is already capable of influencing purchasing behavior.

This distinction changes the investment question itself. Real opportunity hinges on: which transactions become cheaper, easier, or more reliable when software coordinates them, and which companies retain control over distribution, authorization, and fulfillment.

Stablecoins fit into this analytical framework because agents need payment methods. They can streamline digital service procurement and certain cross-border transfers, especially when both parties already hold compatible assets. But agents can also complete payments via bank cards, instant bank transfers, or corporate accounts.

The evolution of product discovery, purchase authorization, and settlement infrastructure each follows independent economic logic and requires its own distinct body of evidence.

The adoption of business models has never depended solely on technology

U.S. Census Bureau quarterly series data documents a gradual expansion interrupted by major shocks. Seasonally adjusted, U.S. online retail share rose from 0.6% in Q4 1999 to 2.2% in Q4 2004, 4.1% in Q4 2009, and 11.2% in Q4 2019.

The average annual growth was 0.35 percentage points in the first decade, accelerating to 0.71 percentage points in the second. The smartphone era coincidentally aligns with faster growth, but this macro sequence cannot disentangle smartphones from improvements in broadband, e-commerce platforms, payment acceptance, or logistics delivery.

Between Q4 2019 and Q2 2020, this share jumped 5.1 percentage points to 16.3%, before falling back to 14.6% in Q4 2021. The forced shift in shopping habits combined with the contraction of offline consumption jointly influenced this ratio.

Data released in August 2026 showed a Q2 2026 share of 17.1%. The U.S. Census Bureau later noted that a data revision originally scheduled for November 19 falls outside this article’s statistical cutoff date.

Figure: Note: Numbered data sources are listed at the end of this article

Due to differing statistical methodologies and denominators, penetration rates across countries cannot be directly ranked horizontally. Europe’s online shopping participation rate rose 16 percentage points over a decade, yet 2025 results still varied from 57% in Bulgaria to 95% in Ireland.

China’s published retail share reveals its online physical goods channel is far larger than Japan’s, while South Korea’s data describes device selection within online commerce. Given China’s adjustments to statistical methodology, constructing continuous growth sequences requires caution, and emerging markets have yet to establish comparable long-term purchase series.

China’s official five-year review offers a more robust comparison: between 2020 and 2025, the share of online retail sales of physical goods in total retail sales of consumer goods rose by 3.3 percentage points, with online sales growing at an annual average rate of 7.6%, explicitly noting that logistics improvements supported this expansion.

Japan’s online share rose 5.03 percentage points between 2015 and 2024, with almost no change in 2025. India’s urban-rural household online shopping gap of up to 23.4 percentage points indicates that nationwide payment coverage alone cannot explain shopping behavior. Income, accessibility to appropriate merchants, and delivery economics continue to separate potential users from actual buyers.

The adoption of payment tools often outpaces the rollout of new retail channels themselves. The Central Bank of Brazil reports over 170 million people have used Pix. RBI data shows 2025–2026 fiscal year UPI transaction volume reached 241.6 billion, both serving broad payment ecosystems well beyond online shopping.

In contrast, a Forrester report shows Apple Pay took over a decade to achieve roughly a 25% three-month usage rate among surveyed adult U.S. internet users. Distribution channels for payment capability, merchant acceptance, and relative advantages over traditional solutions matter far more than launch dates.

Supporting infrastructure has expanded significantly: the World Bank estimates adult account ownership rose from 51% in 2011 to 79% in 2024; the International Telecommunication Union estimates internet usage reached 76% in 2026, with high-income countries at 93% and low-income countries at 32%. Yet neither metric directly measures shopping.

UNCTAD statistics showing approximately $28 trillion in 2024 enterprise ecommerce volume covers 46 economies, including B2B sales.

Historical differences reveal specific mechanisms rather than a universal adoption curve. Reliable delivery and post-purchase recourse make unfamiliar sellers viable, frictionless payments expand the customer base, and dense platform ecosystems enrich supply. Where existing stores are already sufficiently convenient, switching channels yields limited cost savings.

Embedded wallets and super apps eliminate registration steps, while subsidies can accelerate trial but do not necessarily build enduring demand. AI agents inherit much of this infrastructure. Product discovery can roll out as an add-on feature; delegating purchases introduces a higher bar: consumers must trust another decision-maker to manage their funds.

Adoption Evidence Actually Points to Several Distinct Markets

AI-assisted discovery helps identify products; referral tracking logs user clicks from AI services to merchant sites; assisted checkout may still require human confirmation for individual purchases; full delegation relies on pre-authorization, eliminating repeated inquiries per transaction.

Machine-to-machine (M2M) commerce describes counterparties, typically procuring software datasets or inference compute, operating under any of the aforementioned authorization models.

Figure: Consumers utilize multiple AI tools; Source: McKinsey & Company

Model capability benchmarks measure a different frontier. METR's January 2026 recalibration covered 228 software and research tasks, measuring a doubling period for frontier capability time horizons at 131 days since 2023. This time horizon measures human task duration at 50% fit success rate, not production reliability or actual agent runtime.

This result demonstrates rapid model capability gains within evaluated task distributions. Converting this to autonomous retail GMV requires separate evidence proving that procurement performance, merchant access, and customer authorization improve in tandem with model capabilities.

A McKinsey October report draws on June surveys of approximately 21,000 consumers across 11 markets. Its widely cited “53% use AI to shop” figure uses a weighted base of 9,988 individuals, aggregating markets by sample size rather than population.

14% of AI users reported completing purchases directly via AI tools, representing roughly 9% of the broader respondent pool. These rely on self-reported shopping behaviors and do not test autonomy at the transaction level. Even the comparison questions differ: retailer AI “assists” checkout, while platform AI “completes” purchases, making percentages unable to clearly establish preferences across similar services.

Figure: On average, roughly half of consumers now use AI to shop; Source: McKinsey & Company

Figure: Across surveyed markets, Gen Z and Millennials are more likely than older consumers to use AI for shopping; Source: McKinsey & Company

Figure: Consumers are increasingly turning to AI to complete purchases, most commonly in apparel or grocery categories; Source: McKinsey & Company

Actual observed referral data reflects downstream funnel value. In Adobe’s September U.S. retail analytics, AI-referral visitors converted at a rate 90% higher than non-AI traffic, generated 83% more revenue per visit, and referral traffic itself grew 71% year-over-year.

Covering over 1 trillion visits, 100 million items, and 18 categories, the aforementioned conversion rates do not stem from another multi-thousand survey questionnaire.

However, consumers may have completed research before clicking the link, effectively filtering for a highly intent-driven group. Thus, observed conversion advantages do not directly prove AI drives incremental sales, and undisclosed baseline traffic growth masks AI’s true share of retail revenue.

Mercants must distinguish: Is this genuinely an efficient referral channel, or merely capturing purchases that would have occurred anyway via search, ads, or proprietary apps?

Embedded Shopping Shifts the Comparison Baseline

A Reuters November 11 report provides a counterpoint regarding lower external agent transaction volumes. WPIC estimates roughly 5% of Douyin purchases in October originated from chatbot links, exceeding 20% in categories requiring deeper decision-making like nutritional supplements.

JD.com stated nearly 80 million users sought advice from “Dongdong” during Double Eleven, more than triple the year-ago period. These figures represent attribution estimates and company-reported audience sizes, respectively, neither independently disclosing the denominator for autonomous orders.

WPIC’s own promotion analysis describes assistants embedded within Alibaba’s, JD.com’s, and Douyin’s existing shopping environments. This arrangement carries economic advantages: product discovery directly leverages existing catalogs, identity systems, payment methods, and fulfillment relationships, allowing consumers to receive useful assistance without configuring new services.

Evidence from the Chinese market indicates that AI-influenced commerce within platforms has reached substantial scale and should not be dismissed simply because the interaction endpoint remains a manual click.

Similarly, chatbot attribution data cannot reveal how much incremental spending originates from them. A single recommendation might merely replace a search box query, influencer affiliate link, or sponsored ad slot, leaving total purchase volume unchanged. Meaningful experiments should compare purchase volume, margin, and repurchase behavior among demographically similar users exposed to different discovery tools.

Independent external agents face a fundamentally different starting point. During Uber’s November 4 earnings call, management noted roughly 320 million platform trips that Halloween week, with only about 500 Uber Eats orders initiated by agents.

Against this broad denominator, every million combined trips yield only about 1.6 agent-driven food delivery orders, which does not reflect conversion rates among users with access to the feature.

A DoorDash October 29 conference call mentioned hundreds of cumulative third-party agent orders since integration launched. Cumulative counts and weekly counts cannot be linearly annualized, and neither company disclosed qualifying user bases, failed attempt counts, or repurchasing buyer data.

Attitude surveys explain why assisted usage can grow while purchasing delegation lags behind. Gartner surveyed 322 U.S. consumers in May, finding that even in low-risk categories, only 11% were willing to delegate purchasing decisions to AI.

Forrester’s April research found roughly three-quarters of surveyed adult internet users across the U.S., U.K., and Canada feel uneasy about autonomous payments, even with preset limits; their published article did not disclose sample size.

Existing evidence points to two distinct adoption paths. Embedded assistants leverage existing distribution networks while keeping purchasing control within familiar workflows; independent agents must simultaneously demonstrate utility and earn action authorization. Data disclosed along either path so far fails to represent fully delegated commerce across society.

This distinction also affects how experimental data is interpreted. Migrating purchase flows from product pages to chat interfaces can boost convenience without necessarily reducing human involvement; conversely, a small volume of pre-authorized restock orders, despite modest GMV, substantiates genuine delegation.

Ideal disclosures should track a cohort of users through the full journey from touchpoint to purchase, repurchase, and exception handling, flagging precisely when human confirmation was requested. Without this sequence, massive assistant audiences and minimal external order volumes can coexist truthfully, as they describe different stages of the transaction.

Merchant Access and Model Reliability Constrain Delegated Purchases

An agent navigating a checkout interface does not imply granted access permissions. eBay’s Terms of Use explicitly require automation tools to obtain prior consent, expressly covering procurement agents and large language model-powered bots.

A Reuters November 4 report states Amazon blocked Meta’s Muse, while Walmart opted to partner with it. Retailers are competing over who controls transactions, customer data, and attention monetization opportunities.

Authorized APIs can provide accurate product identification, inventory status, and definitive order state; browser automation must infer these states from human-facing UIs and simultaneously navigate merchant security detections.

DataDOME’s October demonstration showed three agents successfully completing purchases in test stores even after being flagged by security systems. This proves technical feasibility in controlled settings but cannot gauge overall retail market acceptance.

Model judgment forms another constraint. Visa’s November 5 benchmark test ran 22 models from 12 labs across 15 shopping scenarios, testing each three times. Average success rate was 57.2%; only 34.2% of model-scenario combinations succeeded all three trials; in failed runs, 54% still ultimately executed purchase decisions.

These scenarios test shopping decisions preceding checkout, payment, and fulfillment, meaning results cannot be directly mapped to real-world payment failure rates.

EComAgentBench offers an alternative diagnostic lens: seven models attempted 662 unit-level tasks in an API sandbox, with requirements distributed across query phrasing, user profiles, and script clarifications. The strongest model achieved 57.1% accuracy; GPT-5 mini recommended items for every task but scored only 31.6% accuracy.

Accuracy here refers to selecting target items or satisfying all requirements, not executing payment or fulfillment. Hidden requirements are harder to capture than explicit ones, making requirement ingestion critical to commercial control: monitoring only task completion rewards throughput while masking incorrect selections.

Extending recorded trajectories from 75 calls to 100 yielded an average accuracy gain of just 0.57 percentage points. Since tests did not log token costs, this only demonstrates limited marginal accuracy gains under evaluated strategies, not economically optimal inference budgets.

Business-grade reliability demands procurement outcomes genuinely satisfy user instructions, not merely generate an authorization code. Selecting the right category but wrong size, subscription cycle, or delivery date can escalate into costly errors. Order flows must also withstand interruptions: if a merchant accepts an order but the agent misses the acknowledgment, automated retries trigger duplicate orders unless both sides can trace the original request.

Liability chains must also pass seamlessly between assistants, merchants, and payment processors. A user’s budget ledger only proves spending authorization, not whether selected items meet actual needs. Refund rights, cancellation windows, and proof of delivery must flow with the order. Introducing human review for unfamiliar merchants or premium substitutes contains error exposure while preserving most coordination gains.

A practical design employs time-bound merchant quotes recording exact specifications, final prices, and delivery terms, followed by a single order submission generating persistent receipts. Major modifications trigger reauthorization or fall outside agent authority entirely.

These controls currently remain advisory operational requirements, not yet standard across existing integrations. But they render performance measurable: price drift, duplicate submissions, and missing receipts can be tracked independently of “recommendation rationality.” Faster checkout only holds commercial value when downstream costs are contained.

Merchant adoption similarly hinges on the economic equation. Agents bringing new customers will welcome them; those merely siphoning existing loyal shoppers to alternate interfaces while charging fees will face resistance. Assessments must factor in contribution margin, displaced ad revenue, referral commissions, support costs, and customer data acquisition. Anonymized browser traffic complicates attribution and accountability further.

Technical Standards Evolve Around Existing Platforms

Different standards are addressing distinct components of this challenge. Google and Shopify helped develop the Universal Commerce Protocol (UCP) to open product discovery and checkout capabilities. The ACP developed by OpenAI and Stripe similarly structures commercial interactions while preserving merchant fulfillment roles and customer relationships.

Current OpenAI documentation heavily emphasizes structured catalogs and product discovery. Integration still rests on merchant willingness, however, as open specifications cannot compel retailers to accept agents.

Purchase authorization requires separate logging mechanisms. Google transferred AP2 to the FIDO Alliance in April, alongside a 0.2 version supporting unattended human transactions. Mastercard’s Verifiable Intent similarly logs authorizations; its March announcement described planned integration into Agent Pay rather than full deployment.

Visa’s Trusted Agent Protocol validates agent identity and ties interactions to consumer and payment details. Stripe’s Shared Payment Tokens provides restricted payment credential access.

These mechanisms address complementary problems: merchants need to identify agents, verify buyer authorization, and process valid credentials. Yet even with signed instructions, product quality, delivery, and post-purchase recourse depend on peripheral commercial agreements.

At the software layer, MCP connects applications to tools and data, while A2A enables agent-to-agent interaction; neither directly moves funds. Resource payment protocols like x402 can operate parallel to these connections, as can legacy billing accounts.

PayPal’s October Agent Ready documentation maps ACP and UCP workflows onto Braintree and PayPal processing systems, demonstrating auditable integration paths without implying universal merchant or regional availability.

A more concrete example is the October deployment by Visa, BILL, and Billtrust, a live pilot using agents to process approved invoice payments through supplier portals. It automates execution within existing authorization frameworks but has not yet published comparative cost or error rate data.

Google began rolling out UCP checkout in Canada on November 4, involving select retailers and certain Shopify merchants. By contrast, DoorDash’s Corporate Dining Connector opened waitlist testing on September 30, reflecting distinct deployment phases between formal promotion and beta access.

Industry-specific customization remains essential. UCP announcements included a Food Committee formed in July (featuring DoorDash, Uber Eats, Google, Square, and Toast) and a draft hospitality capability released in September. Information and operational rules required for restaurant menu substitutions and hotel cancellation policies vastly exceed generic checkout flows.

This commercial strategy aligns with incumbent giants’ logic of protecting core assets: Google supplies traffic discovery, Shopify provides merchant infrastructure, payment processors control payment relationships, and card networks govern acceptance networks and risk frameworks. Open standards reduce the cost of integrating multiple assistants, enhancing merchants’ ability to switch intermediaries while enabling assistants to reach broader supply. Enduring advantages likely remain in distribution and execution quality.

For procurement teams, this means decisions can be decoupled: choosing a commercial interface need not lock a single payment asset, nor does adopting an authorization standard tie you to a specific receiving assistant. Interoperability preserves these options, but actual switching costs still hinge on whether purchase history, consent records, refunds, and negotiated terms migrate seamlessly with integrations. Suppliers resolving these operational dependencies retain revenue even when protocols open.

How Stablecoins Reshape the Economic Equation in Specific Areas

Transfer Volume Does Not Equal Autonomous Commerce

Stablecoins provide software with transferable balances governed by wallet and spending policies, but their economic value depends on what those balances purchase. Total transfer volumes conflate trading, fund sweeps, and affiliated capital flows.

A April BIS speech cited roughly $35 trillion in stablecoin volume for 2025; its 2026 Annual Economic Report relied on differentiated classifications and adjusted datasets yielding ~$28 trillion. Neither figure measures autonomous commerce.

A McKinsey and Artemis foundational payments analysis, based on December 2025 data, estimates payment-related activity annualizes at $390 billion, with B2B payments accounting for $226 billion. The B2B classification assumes 20% of tagged custodial and cash management activities represent payments, cross-validated with transaction size filters.

Treating $390 billion as a full-calendar payment total, or dividing it against annual transfer volumes to calculate penetration, conflates incompatible temporal baselines and definitions.

x402 transaction identification tracks protocol usage, which does not necessarily equate to autonomous procurement.

TRM filtered out self-transfers, highly concentrated payer models, and sellers with fewer than ten independent buyers. Its loose screening targets small, diverse payments; stricter criteria added persistence and evidence of agent registration or multi-merchant presence. Within sampled volumes, USDC comprised 99.6%.

Applying report percentages to the filtered base suggests a potential agent-relevant scale of ~$154k to $1.92m under these assumptions. This represents a categorical bin within observation samples, not a statistical confidence interval or global market estimate. Single-use agents with fixed pricing might fail filtering, whereas routine scripts could mimic agent behavior.

Conversely, evidence argues against categorizing all activity as experimental. Chainalysis reported in October that 30-day wallet top-up retention rose from 70.6% in Q3 2025 to 89.3% in Q2 2026, despite that cohort already excluding wallets transacting for fewer than two days.

Repurchase patterns among specific user cohorts are encouraging, but proving they extend beyond mere wallet retention requires linking these payments to delivered services and non-subsidized vendor revenue.

These datasets vary in observation windows, public chains, and coordinator coverage; their transaction counts and amounts cannot be mechanically summed without deduplication. Commercial validation requires receipts, counterparty identification, and refund tracking. Multiple wallets may belong to a single client, and one account may serve numerous agents, meaning wallet retention does not necessarily mirror client retention.

Circle’s Q3 disclosures further separate financial outcomes from transfer activity: quarterly on-chain USDC volume hit $13.9 trillion, while the Circle Payments Network annualizes at $41.3 billion.

Its revenue and reserve income totaled $865 million, the latter comprising $658 million in reserve returns and $182 million in ARC token presale proceeds.

None of these figures isolate agent payment revenue. For issuers, average balance and reserve yield may matter more than transfer frequency; faster capital velocity actually supports equivalent payment volumes using less idle float.

Software Procurement Offers More Compelling Economics Than Full Retail Replacement

x402 ties server payment requests to resource access, arguing consumers can purchase individual datasets or single inference outputs from unknown vendors without standalone subscriptions. For existing long-term clients of a compute provider, however, pay-as-you-go pricing already serves as a highly competitive alternative.

The Machine Payments Protocol (MPP) from Stripe and Tempo supports stablecoins alongside traditional methods like bank cards implemented via Shared Payment Tokens. Browserbase and Parallel represent typical implementations in browser session and web access domains.

AWS fully launched AgentCore Payments in August, offering wallet integration, spending caps, and observability features. These advancements construct infrastructure, yet recurring commercial revenue matching this capacity remains absent.

The most compelling use cases combine unknown counterparties, micro-procurement, and auto-verifiable delivery. Even with negligible settlement fee variation, bypassing vendor onboarding and avoiding idle subscriptions yields substantial savings; conversely, established relationships usually entail cheaper authentication, credit lines, and reconciliation than freshly topped-up wallets.

Fund portability requires compatible assets, networks, and authorization mechanisms. Reusable balances cut idle collateral across vendors, while shared identities and account-based billing offer alternative pathways. Supplier diversity and access costs dictate the opportunity scope in this space.

Microtransactions Depend on Payment Architecture Design

Stripe’s MPP implementation mandates minimum product thresholds of $0.50 for bank cards and 0.01 USDC for stablecoins. Circle’s Nanopayments documentation outlines pre-funded gateways, off-chain authorized signatures, and batch-delayed settlement mechanics—delivering resources immediately upon verification rather than executing individual on-chain transfers per request.

Assume 100,000 API calls priced at $0.01 each within a month, totaling $1,000, with both parties using USD to avoid FX slippage. For the stablecoin configuration, assume a single settlement cost of $0.001 and aggregated deposit/withdrawal costs of 0.20%.

Add a illustrative $5 monthly operational subsidy to each scenario. These assumptions are not vendor quotes and exclude unmeasured fraud, custody, and credit losses.

For traditional billing, Stripe’s published U.S. standard pricing charges 2.9% + $0.30 for domestic online card payments and 0.8% (capped at $5) for ACH direct debits. Volume is measured monthly with consolidated deduction, avoiding the false comparison of “swiping once per call.”

Under these assumptions, direct transfer network costs must drop below $0.000273 per call to beat monthly card deductions, or fall under $0.000030 to outperform ACH. Even at a $0.0001 rate, direct transfer monthly costs reach $17.

Batch settlement schemes are cheaper due to reduced settlement frequency, a pooling effect traditional prepaid wallets also leverage. Factoring in additional service provider or custody fees could even reverse the comparison.

The relative advantage over ACH amounts to merely $2.90 monthly. Raising aggregated deposit/withdrawal costs from 0.20% to 1% pushes batch scheme costs to $15.10, exceeding ACH but remaining below card payments. Frequent fiat conversion erodes savings; keeping capital within mutually accepted networks dilutes exchange costs, though unusable residual balances introduce offsetting overhead.

These figures encompass processing and operational costs, not just interchange fees. The Kansas City Fed’s comparison segmented jurisdictions, card types, and transaction classes; St. Louis Fed analyses measured bank interchange revenue rather than full merchant costs. Gas sponsorship merely shifts costs; service fees, top-ups, compliance, reconciliation, and loss sharing remain material.

The table above aggregates total costs borne by each party: top-up fees may fall on buyers, gas on coordinators, reconciliation split between counterparts. The flat $5 subsidy isolates processing stage differences rather than simulating real workloads. Production-grade comparisons require contract quotes, empirically observed losses, and adoption-side allocation ratios, because even if total costs decline, individual adopters may still suffer worse positioning.

Stripe separately prices stablecoin acceptance at 1.5%, bundling fiat conversion, compliance screening, and gas sponsorship. This packaged product differs fundamentally from the settlement configuration assumed above: cheap blockchain transfers alone cannot anchor comprehensive merchant service pricing.

Compare Against Complete Transaction Lifecycles

Preloading $1,000 at month-start with uniform monthly spend yields an average idle balance of $500 over 30 days. At an 8% annualized financing cost, capital costs amount to $3.29 regardless of whether funded via card or stablecoin. Conversely, suppliers awaiting $1,000 invoice repayment over 30 days incur ~$6.58 in financing costs at equivalent rates.

Instant payments eliminate accounts receivable but accelerate buyer cash outflows. These examples rely on divergent balance schedules and cannot simply offset each other.

Batch settlement reduces transaction fees but leaves sellers exposed prior to final clearance. Prepaid models restrict default risk, while shorter settlement cycles and spending caps help contain outstanding claims, though verification, reconciliation, and settlement costs persist.

Cross-border savings hinge on origin and destination funding. USD stablecoins bypass intermediate currency conversions and banking hour constraints, yet local fiat deposits and withdrawals still carry exchange spreads and ancillary fees. Indigenous alternatives are inherently competitive: Eurozone instant payment rules expanded coverage in October 2025, while Pix and UPI deliver mature real-time infrastructure.

Traditional payment networks are absorbing stablecoins. Visa reported September stablecoin annualized settlement volume surpassed $20 billion. Consumers keep swiping cards while institutions settle behind the scenes with stablecoins, leaving the entire checkout experience untouched.

Finally, faster settlement scarcely resolves procurement errors. Disabling chargebacks shifts greater responsibility onto verification, refunds, and buyer protection. API responses may validate instantly, but hotel bookings require safeguards extending until check-in. Blockchain confirmations establish neither legal finality nor guaranteed convertibility into bank deposits; a mistaken purchase remains erroneous regardless of settlement speed.

Regulation Remains an Embedded Cost

The European Securities and Markets Authority (ESMA) issued a statement on October 8 directing licensed crypto-asset service providers to halt non-compliant (MiCA-ineligible) stablecoin offerings to EU clients, permitting only limited remedial arrangements.

The European Banking Authority (EBA) designated USDC as a significant e-money token on October 14. Regulatory handover was originally scheduled for November 12, falling outside this article’s statistical cutoff.

The Office of the Comptroller of the Currency (OCC) released final GENIUS Act rules on October 21 outlining reserve, redemption, custody, and regulatory mandates for regulated entities, effective January 18, 2027, with anti-money laundering and sanctions provisions awaiting separate rulemaking. These timelines dictate available counterparties and compliance expenditures.

Regardless of settlement mode adopted, agents still require enforceable budgets, whitelisted vendors, time-bound authorizations, and auditable receipts.

Which Transactions Merit Delegation to Agents?

The economic litmus test lies in whether savings on time, pricing, and coordination exceed expected losses from reasoning, integration, regulatory oversight, and errors. High-frequency repetitive procurement magnifies trivial per-unit savings, yet optimizing a routine reorder taking two minutes pale compared to reconciling invoices across disparate vendor portals.

Enterprise procurement operates under established approval hierarchies; agent value emerges in handling exceptions or expanding vendor coverage beyond current automation boundaries.

At home, narrow-scope restock directives can strictly define items, quantities, and price ceilings; team dining entails coordinating multiple preferences under a unified budget; complex travel promises massive savings but risks astronomical penalties if cancellation policies or connection timings fail. The strongest initial deployment scenarios typically combine “restricted authorization” with “verifiable outcomes before loss accumulation.”

Jane Street’s Is the AI Singularity Incoming? roundtable emphasized short feedback loops, providing a sharper litmus test: what evidences successful purchasing, when do proofs arrive, and which steps remain reversible? Valid API responses confirm delivery, not accuracy; SKU verifications confirm model numbers, not fit or satisfaction.

We infer that delegation becomes attractive only when verification costs and residual expected losses fall below coordination savings. Contracts specifying observable acceptance criteria and recourse rights expand this applicability; in domains marked by subjective outcomes or delayed feedback, stronger model capabilities may still fail to eliminate expensive manual checks.

Value Creation and Value Capture May Diverge

Controlling product discovery is inherently lucrative. Amazon reported Q3 advertising revenue of $19.637 billion, up 11% year-over-year, although this figure transcends retail-sponsored listings. Reuters noted in November that both Amazon and Walmart are testing ad injections within their respective shopping assistants.

Amazon stated agentic shopping has yet to materially impact its advertising business, though this landscape could shift as more purchasing decisions migrate to AI interfaces.

A Wharton School study conducted in August 2026 provided early clues regarding how these decisions might be influenced. Researchers ran approximately 26,000 simulated sport watch selections across six models and eight products.

In a separate experiment involving four models, presenting three recommendation sources simultaneously pushed GPT-5.5’s selection of Wirecutter-recommended products up by 53 percentage points versus baseline, whereas sequential presentation of identical sources yielded negligible effects, with the underlying merchandise completely unchanged.

Though limited to simulated selections rather than real transactions, findings suggest whoever controls information routing to assistants shapes merchant-received demand.

Independent assistants can leverage aggregated purchasing intent to negotiate lower prices or higher commissions. Retailers respond by embedding analogous tools within existing platforms, where they already command catalogs, customer relationships, and fulfillment.

Standardized interfaces may depreciate baseline product-match value, yet inventory depth, distribution networks, and reliable dispute resolution remain structurally harder to replicate.

For merchants, economics pivot on whether agents generate genuinely new customers. Assuming a 20% contribution margin and 5% commission on all agent-referral sales, pure incremental volume must exceed 25% of referred revenue to justify payouts, given baseline buyers acquire without marginal cost.

Reduced ad spend improves this calculus, while returns, support costs, and lost customer data drag profitability. Judging viability solely by agent-generated GMV proves nearly impossible.

Customer relationships outweigh isolated transactions. Assistants retaining purchase histories and preferences can reroute future orders toward competitors, diluting merchant customer lifetime value. Open product data streaming lets merchants reach multiple assistants, mitigating reliance on any single intermediary.

Regardless of underlying settlement method, payment processors profit from credential management, fraud prevention, and reconciliation. Stablecoin issuers benefit from retained float, while wallet providers and payment infrastructure firms require sufficient fee revenue to cover operating costs. Rising transaction volumes do not automatically translate to retained revenue growth.

Commission-based recommendation engines introduce further complexity: assistants might favor paying merchants, skim negotiation discounts, or route orders to the highest bidder. While none represent current market defaults, each potentially pits buyer interests against intermediary revenue.

Evaluating these models demands transparency around compensation mechanics, final transaction prices, and repurchase behavior. The core question extends beyond how many transactions agents facilitate, asking instead how much value they create and who ultimately captures it.

Conclusion

The fastest-moving domain for agentic commerce involves AI-enhanced existing purchase relationships. Embedded assistants leverage native catalogs, payment credentials, and fulfillment networks to effectively influence discovery and conversion.

Fully delegated purchasing introduces elevated barriers: consumer authorization, merchant access, and confidence that outcomes satisfy user directives. Consequently, assisting feature growth creates avenues for delegation but serves as an unreliable adoption proxy.

Economic boundaries center on verifying the cost of confirming agents execute correctly. Narrow-scope procurements featuring observable outcomes and actionable remedies offer clearest entry points; in domains marked by subjective preferences, delayed deliveries, or expensive trial-and-error, verification and oversight expenses may devour saved time. Expanding model capabilities widen feasible market boundaries, while peripheral commercial controls determine execution fidelity.

Distribution channels remain central to value capture. Purchase experiments reviewed herein demonstrate item selection hinges on which information reaches assistants and how it surfaces. Even with open checkout protocols, input control dictates merchant demand. Sustained profitability from this influence depends on incremental sales, customer retention, and intermediaries’ ability to charge without harming buyer or seller economics.

When stablecoins reduce vendor onboarding or settlement friction, they reinforce this commercial thesis. Yet economic viability remains tethered to deposits/withdrawals, financing costs, merchant acceptance, and post-purchase recourse; traditional billing systems persist as formidable competitive alternatives.

Success or failure in the next phase must be judged by paid repurchase rates, marginal merchant contributions, and total cost per correctly executed procurement. Providers rendering delegation measurable, economically sound, and correction-capable will secure stronger retained revenue foundations as adoption accelerates.