Skip to content
Primores AI wiki

Agentic Commerce — The $1 Trillion Shift

Agentic Commerce

TL;DR: Agentic commerce — shopping powered by AI agents acting on behalf of consumers — could represent $1 trillion in US B2C retail by 2030 ($3-5 trillion globally). It’s not just e-commerce evolution; it’s a rethinking of shopping itself where AI anticipates needs, compares options, negotiates, and executes transactions.

What Is Agentic Commerce?

In traditional commerce, you navigate websites, compare products, and make decisions yourself. In the agentic era, AI agents do this for you:

  • Anticipate your needs (calendar shows a move → agent starts planning)
  • Navigate options across multiple platforms
  • Negotiate deals and bundle purchases
  • Execute transactions — sometimes without human involvement

“We’re entering an era where AI agents won’t just assist—they’ll decide. Business models need to evolve from optimizing clicks to earning trust from algorithms acting for consumers.” — Naveen Sastry, McKinsey Senior Partner

Market Projections

MetricValueSource
US B2C retail by 2030$1 trillionMcKinsey
Global projections$3-5 trillionMcKinsey
AI agent involvement by end 202625-30% of US online purchasesIndustry estimates
AI agent involvement by 202850% of online purchasesProjections

Three Interaction Models

1. Agent to Site

AI agents browse merchant websites, extract product data, and make recommendations.

Implication: Your site must be agent-readable, not just human-readable.

2. Agent to Agent

Consumer’s AI agent negotiates with merchant’s AI agent.

Example: Your agent negotiates a hotel upgrade with the hotel’s pricing agent.

3. Brokered Agent to Site

Platform agents (ChatGPT, Perplexity) intermediate between consumers and merchants.

Implication: You may never see the customer — only their agent.

Six Domains Merchants Must Address

To thrive in the agentic era, businesses must adapt across six key domains:

Innovation Required (Build New)

DomainWhat to Do
Customer Engagement & DiscoveryDevelop agents that understand intent and suggest products. Embed semantic metadata in catalogs.
Clienteling & LoyaltyBuild persistent customer-context layers accessible by agents. Expose loyalty APIs.

Renovation Required (Upgrade Existing)

DomainWhat to Do
Core Commerce PlatformsEnable agents to execute structured transactions with minimal human input. Add dynamic pricing, inventory-aware recommendations.
Payments & FraudDifferentiate benign agents from malicious bots. Verify intent and identity in real time.
In-Store Point of ServiceSync digital and physical journeys. Integrate spatial computing for in-store agent navigation.
Fulfillment & ReturnsAgent-ready fulfillment APIs. Automated return logic negotiation.

New Revenue Models

Traditional ad revenue will decline as agents bypass ads entirely. New models emerging:

ModelHow It Works
Multibrand BundlingAgent bundles purchases across brands; each gets revenue share
Real-time Negotiation FeesAgent negotiates upgrades/deals; platform takes success fee
Premium Agent SkillsSubscription access to specialized agents (fashion stylist, trip planner)
Data Insights SalesBrands pay for anonymized agent-filtered behavior analytics
Conversational MarketplacesPurchases via dialogue; monetize via listing fees and commissions
Interagent Protocol FeesFees for cross-platform agent interoperability
Sponsored Smart SuggestionsSubtle, intent-aligned suggestions (preserves user trust)

The Trust Challenge

“When a person walks into a store, the trust equation is straightforward: Do I trust this brand, this merchant, this product? When an AI agent shops on your behalf, trust becomes abstract, filtered through layers of data, automation, and institutional frameworks.” — McKinsey

Trust Is Contextual

What works in one market fails in another:

  • Germany/Japan: Still prefer traditional payment methods, account-to-account transfers
  • These markets may resist delegating purchase decisions to AI longer

Trust Must Be Earned Through Interaction

  • Clear communication, not just legal disclaimers
  • Users must define boundaries of trust
  • Consent must be a living, flexible agreement — not a checkbox

Three Risk Categories

1. Systemic Risk (The Snowball Effect)

When agents are interconnected, minor errors have exponential impact:

  • A faulty prompt triggers cascading unintended consequences
  • Incorrectly booked flights, overordered inventory, unauthorized purchases

Question: Do your agents fail gracefully? Can they backtrack?

When an AI agent makes a poor decision, who’s to blame?

  • The platform that developed the model?
  • The brand that deployed the agent?
  • The user who approved it?

EU AI Act provides some clarity; US regulations remain fragmented.

3. Data Sovereignty (Geopolitical Challenge)

  • Where is agent data processed?
  • Which country’s laws apply?
  • Data localization requirements vary globally

The infrastructure layer (April–May 2026 update)

When this page was originally written (April 2026), agentic commerce was a category projection. Between September 2025 and May 2026 the production infrastructure shipped. The category is now operationally real.

Four standards now constitute the production protocol stack (covered in depth in glossary/agent-payment-protocols):

LayerProtocolLeadLaunched
Commerce schemaUCP (Universal Commerce Protocol)Google + Shopify + retailer coalition (Etsy, Wayfair, Target, Walmart, endorsed by Visa, Mastercard, Stripe, Best Buy)Coalition formed early 2026
Identity / verificationVisa TAP (Trusted Agent Protocol)Visa + CloudflareOctober 14, 2025
Payment authorizationAP2 (Agent Payments Protocol)Google + 60+ partnersSeptember 16, 2025
Payment settlementx402Coinbase + Cloudflare (x402 Foundation)September 2025

The protocols compose — a typical transaction uses UCP for discovery, TAP for agent verification, AP2 for authorization, x402 for stablecoin settlement (or traditional rails). See glossary/agent-payment-protocols for the full architecture.

Production scale (March 2026): x402 alone processed 119M transactions on Base + 35M on Solana, ~$600M annualized volume, ~2-second settlement, zero protocol fees. Stablecoin micropayments at this volume make agent-mediated commerce economically viable in ways credit-card rails could not (interchange fees made penny-scale transactions impossible).

AWS Bedrock AgentCore Payments shipped May 7, 2026. This is the inflection point for enterprise adoption. Agents built on Amazon Bedrock can now discover x402-gated endpoints, settle stablecoin payments, and continue executing — within session-level spending limits and policy-based controls — without human approval per transaction. The Coinbase x402 Bazaar via AgentCore Gateway exposes 10,000+ x402 endpoints discoverable by agents at runtime.

The closed-loop vs. open-protocol dynamic resolved into two viable approaches:

ApproachClosed-loop (Amazon)Open-protocol (Google coalition)
Merchant of recordAmazonThe retailer
Customer loyalty dataAmazon owns itRetailer retains it
Discovery surfaceInside AmazonOpen web
SettlementAmazon internalAP2 + any rail (including x402)

OpenAI’s Instant Checkout shut down March 2026 — the third-position aggregator-with-own-checkout approach lost. The market consolidated to two viable approaches: own everything (Amazon) or be one piece of an open protocol stack (Google coalition).

For brand strategy (marketing/preparing-for-agentic-ai): the architectural choice — Amazon closed-loop vs. open-protocol participation — is now a strategic question with concrete implications for customer-data ownership and merchant economics.

The agent quality gap (Anthropic Project Deal, April 2026)

A finding worth flagging because it shapes how agentic commerce affects consumers asymmetrically. Anthropic ran a real internal marketplace experiment for one week in December 2025: 69 employees × $100 budget each, all transactions through agent representatives. Four parallel marketplaces ran simultaneously — two with everyone using Claude Opus 4.5; two with 50/50 randomization between Opus 4.5 and Haiku 4.5 (less capable model).

Results: 186 deals, >$4,000 in total value in one week. Users represented by more capable models got objectively better outcomes — better prices, faster deals, more favorable terms. Users on the losing end didn’t notice. Anthropic explicitly named this the “agent quality gap.”

Implication for agentic commerce: the quality of the agent representing each consumer becomes a material economic variable. Users with access to better models compound advantage over users with worse models — in ways the worse-served users won’t detect. This is a different inequality vector than traditional pricing or access, and it doesn’t show up in the consumer’s experience as “I’m being treated unfairly” — it shows up as “this seems fine.”

Connection to glossary/agent-adoption-frictions and glossary/ai-agent-behavior: agent biases were the May 2025 finding (what agents choose); agent quality gaps are the April 2026 finding (how well agents represent users at the same task). Both are now load-bearing for agent-mediated commerce.

US consumers currently have no clear dispute rights when AI agents transact on their behalf. The Fair Credit Billing Act, state UCC variants, and chargeback regimes were all written assuming the consumer initiates the transaction. When an autonomous agent initiates a transaction within session-level limits the consumer pre-authorized:

  • Is the agent the consumer? (Legally, probably not.)
  • Is the consumer responsible for what the agent does within authorized limits? (Probably yes, but no case law confirms.)
  • What’s the dispute process when the agent buys the wrong thing, or from a fraudulent merchant? (No standard answer.)
  • Who’s liable when a merchant is designed specifically to deceive AI agents? (No regulatory framework.)

The 2026 protocols (AP2, TAP) include technical structure for dispute flows, but the legal scaffolding hasn’t caught up. Expect substantial regulatory development over 2026–2027 as volume scales past the threshold where individual disputes start producing court cases.

Strategic Questions for Leaders

First-Mover Advantage:

  • How can your business build a defensible moat through strategic API development?
  • What tech infrastructure and partnership ecosystem do you need?

Concierge Experience:

  • How do you create a unique AI-powered concierge that drives loyalty?
  • What can your brand agent do that generic agents cannot?

Revenue Protection:

  • As AI disintermediates ad revenue, what new models can you create?
  • What data monetization or subscription models make sense?

Consumer Trust:

  • How do you earn trust when delegating decisions to autonomous agents?
  • What transparency and human-override features do you need?

AI Agent Bias Factors

Recent academic research (Columbia + Yale, August 2025) reveals that AI agents have predictable biases in purchasing decisions — and these can be influenced.

What Influences AI Agent Product Selection

FactorDirectionNotes
Keyword order in title✅ CriticalMatching search terms exactly: +80pp selection increase
Ratings✅ Positive+0.1 rating increase improves chances
Review count✅ PositiveMore reviews signal trustworthiness
Positive badges✅ Positive”Bestseller”, “Recommended”, “Our Pick"
"Sponsored” label❌ NegativeReduces selection probability

Real-World Example

ZDNet tested ChatGPT’s buying agent for a housewarming present. BlancPottery was chosen because:

  • Tags and badges like “Etsy Recommended”
  • 5-star rating with several reviews
  • Keywords matching search: “Dinnerware Set”, “Handmade”

Model Differences Matter

Different AI agents weight factors differently:

  • GPT-4.1 preferred top-left positioned products
  • GPT-5.1 showed the opposite preference
  • Claude, Gemini, and GPT all have unique bias profiles

Implication for merchants: Test product pages across multiple AI agents, and re-test after major model updates.

Improving Decision Quality

AI agents are getting better at objective decisions. When presented with identical products where one had a 1% discount (objectively better):

ModelFailure Rate
Claude Opus 4.54.3% (down from 63.7% in Sonnet 3.5)
GPT-5.11% (down from 25.8% in GPT-4o)
Gemini 2.5 Flash0% (down from 2.8% in 2.0 Flash)

Key insight: As models improve, gaming tactics become less effective. Focus on genuine value and clarity.

Merchant Recommendations

  1. Optimize for AI agents, not only humans
  2. Understand which AI agents your customers use most
  3. Test different agents and adjust product pages
  4. Re-test after model updates — decisions change drastically

See seo/agentic-search-optimization for detailed optimization tactics.

Key Quote

“This is not a wait-and-see moment. Before long, nearly all retailers will have to grapple with the fact that a significant percentage of their customers will not be human users but rather AI agents. The companies that move first will be the ones that help shape the future.” — Lareina Yee, McKinsey Senior Partner

Key Takeaways

  • $1 trillion US market by 2030; this is not speculative
  • Three interaction models: agent-to-site, agent-to-agent, brokered
  • Six business domains require innovation or renovation
  • Ad revenue will decline; new monetization models required
  • Trust is foundational infrastructure, not just sentiment
  • First-movers will define the standards

Human Psychology vs. Agent Logic

Today’s social commerce platforms (TikTok, Instagram) exploit human psychological triggers:

  • Personalized recommendations → emotional arousal
  • Social proof (likes, reviews) → trust and FOMO
  • Scarcity cues → urgency and impulse action

These triggers compress human decision-making and drive impulse purchases. See marketing/social-commerce-psychology for practical application.

The agentic shift: When AI agents shop, these triggers may work differently:

  • Agents can verify scarcity claims via inventory APIs (fake urgency won’t work)
  • Social proof may become “agent proof” — what other AI agents recommend
  • Personalization becomes even more precise with full purchase history access
  • FOMO doesn’t affect algorithms the same way

Key question: Will merchants need separate optimization strategies for human buyers vs. AI agents?

  • seo/agentic-search — How AI agents decide which brands get found
  • seo/agentic-search-optimization — The ASO discipline
  • marketing/preparing-for-agentic-ai — Brand strategy for agentic era
  • marketing/social-commerce-psychology — Psychological triggers that drive human purchases
  • glossary/ai-agent — What AI agents are
  • glossary/ai-agent-behavior — Agent-side biases that decide what agents choose
  • glossary/agent-adoption-frictions — User-side barriers that decide whether agents get used at all (the demand-side counterpart to all the supply-side optimization in this page)
  • glossary/cohort-analysis — AI agents as customers complicate cohort analysis: the cohort definition changes when the “customer” is an algorithm. The capital-efficiency layer needs to adapt to agent-mediated purchasing
  • glossary/agent-payment-protocols — The four-protocol infrastructure stack (AP2 / x402 / UCP / Visa TAP) + AWS Bedrock AgentCore Payments + the agent quality gap finding from Anthropic Project Deal. The production substrate this page operates on top of.
  • tools/gemini-omni — May 2026 launch of Google’s any-to-any multimodal model adds a video-generation layer to the agentic-commerce stack, paired with Gemini Spark (24/7 personal agent) and Antigravity 2.0. The agent-platform API is the surface where commerce agents and multimodal generation will meet.

Sources