Agentic commerce is buying conducted by software agents, from both directions: on the customer’s side, AI agents discover, evaluate, and increasingly purchase on the customer’s behalf; on the merchant’s side, autonomous agents adjust pricing, merchandising, promotions, and post-purchase engagement in real time. The merchant side optimizes the storefront that exists today. The buyer side replaces the shopping journey itself — and it is the less understood, faster-moving half.
On the merchant side, the shift is driven by the limits of traditional e-commerce optimization. Human merchandisers can manage dozens of product categories and a handful of promotional strategies. AI agents can optimize across millions of product-customer combinations, thousands of pricing scenarios, and dozens of channels — continuously and in real time. Early adopters like Amazon have used algorithmic commerce for years; agentic commerce extends this to mid-market retailers and brands through accessible AI agent frameworks and data platforms that give agents the unified customer context they need.
Agentic commerce connects closely to digital commerce but adds a layer of autonomous intelligence. Where digital commerce describes the channels and infrastructure for online selling, agentic commerce describes the AI-driven operating model layered on top of them. It is the commerce-side counterpart to agentic marketing, applying the same autonomous plan-execute-learn loop to the storefront rather than campaigns.
When the agent becomes the buyer
Agentic commerce has two sides, and they are not equally understood. The merchant side — dynamic pricing, personalized merchandising, checkout optimization — is an optimization story. The buyer side is the structural change: the customer’s agent, not the customer, runs discovery, comparison, and increasingly the checkout and payment.
The buyer side is where the evidence is accumulating. Traffic to US retail sites from generative AI tools rose 693.4% year over year during the 2025 holiday season (Adobe Analytics, 2026-01-07), and Adobe’s analysis of the same season found AI-referred visitors converted 31% more often than other traffic sources — from a base that remains modest, visible today mainly as pilots: agent checkout inside ChatGPT on the OpenAI–Stripe protocol below, plus retailer experiments, rather than scaled channels. Half of all consumers already use AI when searching the internet (McKinsey, 2025-10-17). And delegation has begun: 24% of consumers say they are already comfortable with an AI agent shopping for them, rising to 32% of Gen Z (Salesforce, published 2024-10-29). McKinsey sizes the prize at $900 billion to $1 trillion in agent-orchestrated US B2C retail revenue by 2030, and $3 trillion to $5 trillion globally (McKinsey, 2025-10-17).
| Stage | Human-led buying | Agent-mediated buying |
|---|---|---|
| Discovery | Shopper searches, opens tabs, compares options | Agent queries catalogs and returns a ranked shortlist |
| Evaluation | Shopper reads reviews and product pages | Agent weighs reviews, specs, prices, and policies as data |
| Basket | Shopper assembles and edits the cart | Agent assembles the cart against the customer’s stated constraints |
| Checkout | Shopper fills forms, re-enters payment details | Agent calls merchant checkout endpoints directly |
| Payment | Shopper authorizes the transaction | Agent pays under a verifiable mandate from the customer |
| After the sale | Shopper tracks the order, starts returns | Agent monitors delivery and initiates returns within policy |
From pages to machine-readable product truth
When agents do the buying, the merchant’s surface area moves from pages to data. A human shopper tolerates a stale shipping table or a returns policy buried in an FAQ; an agent parses price, availability, and policy as fields and acts on them at face value. Product data, price, inventory, delivery promises, and the terms that surround the transaction — returns, warranties, price-match rules — become the interface the agent reads. Agents parse; humans browse. The merchant has to serve both from the same truth.
McKinsey frames the requirement the same way: winning merchants will embed semantic and behavioral metadata in product catalogs and expose agent-authenticated interfaces, rather than assuming a human is looking at a rendered page (McKinsey, 2025-10-17). The merchant protocols make the point concrete. Under the Agentic Commerce Protocol, the open standard developed by OpenAI and Stripe, a merchant’s checkout session response must return authoritative cart state — items, pricing, taxes, fees, shipping, discounts, totals, and status (ACP specification, 2026-04-17). The merchant’s system, not the agent’s memory of a page, is the source of truth for what the order contains and what it costs. The same 2026-04-17 snapshot added a product feed specification, extending the machine-readable surface backward from checkout state to the catalog itself — the feed is what lets an agent compare availability and price before a cart ever exists.
Customer context: why a CDP plus an AI agent is the minimum viable stack
An agent can only buy well for a person if it knows the person. Preferences, purchase history, sizes, household constraints, and consent all travel with the buyer — and on the merchant side, the same context decides whether the transaction is intelligent or generic. McKinsey describes the target as persistent customer-context layers that agents can access through APIs (McKinsey, 2025-10-17). A customer data platform is that layer: identity resolution stitches the customer’s records into one profile, and the profile is served in real time, at the moment the agent assembles a cart.
Consent is part of the context, not a footnote. What the customer agreed to — which channels, which uses, which data — has to be a field the agent can read, which is why consent management belongs inside the commerce stack rather than beside it. Without unified context, agent decisions degrade to the data in front of them: a pricing agent discounts a loyal full-price buyer because it sees only the transaction, a discovery agent ranks results from browsing history alone, a post-purchase agent sends replenishment messages on a fixed schedule instead of predicted need.
This is the merchant-side case for an Agentic CDP: a platform that combines data unification, predictive analytics, and native activation so the agent that serves the customer reads the same profile the marketer does. CDP plus AI agent is the minimum viable stack for the merchant side of agentic commerce — the agent is the decision-maker, the CDP is the memory.
The rails: agent-readable checkout, payment mandates, disputes
Deals need rails. The question every merchant must answer is narrow and unforgiving: when an agent assembles the cart, who authorizes the payment? Three structures now exist.
Agentic Commerce Protocol (ACP). An open, Apache 2.0-licensed standard maintained by OpenAI and Stripe — pre-stable overall, with its latest dated snapshot, 2026-04-17, labeled stable. The merchant implements create, update, and complete checkout-session endpoints; the agent assembles the order inside surfaces such as ChatGPT, and the merchant processes payment on its own rails. A separate Delegated Payments specification lets credentials travel as tokens — Stripe’s Shared Payment Token is the first implementation (agenticcommerce.dev; github.com/agentic-commerce-protocol).
Agent Payments Protocol (AP2). Introduced by Google in September 2025 as an extension of the Agent2Agent protocol and Universal Commerce Protocol (McKinsey, 2025-10-17; ap2-protocol.org). AP2 answers the authorization question with mandates: a user-signed Checkout Mandate — bound by hash to the merchant-signed checkout, or constrained by user-signed open mandates in human-not-present flows — gives the merchant cryptographic proof the agent was authorized to buy that cart, and a Payment Mandate gives the credential provider, network, and processor proof the agent was authorized to pay. It defines two modes — human present, where the customer approves the exact cart, and human not present, where the customer sets constraints and the agent acts inside them (AP2 specification, v0.2).
Network-level agent credentials. Mastercard Agent Pay issues payment credentials built for agents, restricts transacting to registered agents (“know your agent”), and requires verified user intent and explicit consent before an action is taken (mastercard.com, Agent Pay).
Disputes are part of the rail, not an afterthought. AP2’s mandates chain into a non-repudiable audit trail explicitly designed to support dispute resolution; the merchant side still needs a refund and return workflow that accepts agent-initiated requests within policy, and an order record that links mandate, checkout, payment, and fulfillment.
Trust becomes the conversion bottleneck
In agent-mediated commerce, trust stops being a brand sentiment and becomes a conversion variable: an agent that cannot verify who it is dealing with, or whether the data is honest, stalls the transaction. Three layers decide it.
Verified identity. Both sides of the transaction must be provable — the agent’s authority to buy (mandates, registered-agent checks) and the merchant’s legitimacy behind the endpoint. Mastercard’s registered-agent requirement and AP2’s verifiable credentials are the emerging mechanisms.
Brand consistency. The data the agent reads must match the experience the customer receives. A price, delivery date, or return policy that differs between the feed and the fulfillment system breaks the promise the agent already made on the merchant’s behalf — and the customer blames the merchant, not the agent.
Review integrity. Agents weigh reviews, ratings, and return statistics as data. Fabricated or incentivized reviews distort agent decisions at machine scale, because the agent cannot browse its way around them the way a skeptical human might. Merchants who invest in honest review signals are, in effect, investing in how their catalog reads inside an agent.
The consumer baseline is impatient: 72% of consumers trust companies less than a year ago and 65% say companies are reckless with customer data (Salesforce, surveyed July 16–August 16, 2024, published 2024-10-31). Nearly 75% want to know when they are dealing with an agent rather than a person, and 45% say a clear escalation path makes them more likely to use one (Salesforce, 2024-10-31). A shopping agent that represents a brand inherits that skepticism — the merchant’s transparency obligations travel with it. The consumer-facing surface of this shift is the AI shopping assistant, and the same trust rules apply in both directions.
Governance and guardrails for autonomous agents
Autonomy is the point of agentic commerce, and it is also where programs fail. An agent left to optimize a single metric learns sharp-edged behavior: stacking discounts until margin disappears, showing manufactured scarcity, or pushing high-margin products over what the customer actually needs. Governance is not a compliance afterthought — it is the mechanism that lets autonomy expand without those outcomes.
Effective programs tier decisions by reversibility and customer impact. Cheap, reversible decisions such as search ranking and recommendation order can run autonomously under monitoring. Expensive or hard-to-reverse decisions — price changes above a threshold, refunds, subscription cancellations — need human approval or strict numeric limits. When several agents act on the same customer, AI agent orchestration defines which agent owns which decision and resolves conflicts, so the pricing agent’s discount does not fight the merchandising agent’s margin target.
| Agent decision | Autonomy level | Guardrail to establish | Failure mode if ungoverned |
|---|---|---|---|
| Search ranking and recommendations | Fully autonomous | Quality monitoring, result-diversity checks | Feedback loops narrow what customers ever see |
| Discounts and promotions | Bounded autonomy | Margin floors, budget caps, eligibility rules | Margin erosion; customers trained to wait for discounts |
| Price changes | Human approval above thresholds | Change ceilings and an approval workflow | Pricing errors published to every channel at once |
| Checkout interventions | Bounded autonomy | Frequency caps per customer and session | Nagging that suppresses the conversion it chases |
| Refunds and cancellations | Human escalation | Monetary thresholds and escalation paths | Compounding financial exposure |
Two failure modes deserve names. Objective drift: agents optimize the metric you gave them, not the one you meant — a conversion-maximizing agent will trade margin for a fraction of a point unless margin sits inside the objective or the guardrails. Bias reinforcement: agents trained on past behavior reproduce its distortions, so AI bias in marketing reviews belong in the operating cadence rather than in a one-time audit. The remedy for both is the same — log every agent action with the context it saw, review a sample on a fixed schedule, and be willing to claw back autonomy from a tier that misbehaves.
Measuring agentic commerce
Agentic commerce changes what “working” means, so it needs its own scorecard. Conversion rate alone cannot tell you whether agents create value or quietly pay for it with discounts, incentives, and free shipping. A measurement plan has to track the trade the agent is making, not just the outcome it was told to maximize.
Four measures cover most programs:
- Autonomous share of decisions — the percentage of pricing, merchandising, and intervention decisions agents made without human approval. If it never rises, the program is automation theater; if it rises while the other measures hold, autonomy is earning trust.
- Margin retention — revenue minus the discounts and incentives agents issued. A conversion lift paid for by margin giveback is not a win.
- Guardrail and escalation events — how often agents hit limits or required human review, and why. Rising events are an early warning; a silent log usually means monitoring is blind.
- Experience quality — opt-outs, complaints, and return rates show whether adaptive experiences serve customers or exhaust them, a concern the agentic customer experience discipline treats in depth.
Instrumentation comes before dashboards: every agent action needs a record of the context it saw, the action it took, and the outcome that followed, or none of these measures can be computed. Holdout groups remain the honest referee — keep a small segment on the pre-agent experience and compare incrementally, because continuously optimizing agents will otherwise claim credit for seasonal effects they did not cause. Agent-referred traffic deserves the same discipline: segment it, track its conversion and margin separately, and do not let it disappear into blended totals.
An adoption path for agentic commerce
Agentic commerce is an application of agentic AI, and it inherits that field’s central lesson: autonomy is granted, not switched on. Retailers that hand agents the whole storefront on day one have no baseline, no calibrated trust, and no way to tell a good agent from a lucky week. A phased path trades some speed for evidence at every step.
| Phase | What the agent controls | Prerequisite to advance | Typical failure mode |
|---|---|---|---|
| 1. Observe | Nothing — agents recommend, humans execute | Logged recommendations with predicted impact | Quiet shelving: recommendations pile up unread |
| 2. Assist | One surface, humans approve each action | Approval latency low enough to matter | Bottleneck: humans become the rate limiter |
| 3. Act with bounds | One surface autonomously inside numeric limits | Guardrail breach monitoring in place | Limits set so wide they never trigger |
| 4. Compound | Multiple surfaces on one shared customer profile | Cross-agent conflict resolution working | Agents optimizing against each other |
Two practices keep the path honest. Advance on evidence, not on the calendar — a phase ends when its failure mode is under control, which is what the prerequisite column encodes. And treat the first autonomous domain as a probe: search ranking and recommendations are the usual choice because mistakes there are reversible and visible in seconds, while pricing and checkout autonomy come later, once guardrail telemetry has earned organizational trust.
A merchant readiness checklist for agentic commerce
- Machine-readable catalog. Price, availability, delivery promise, returns policy, and warranty terms exist as structured fields, refreshed at transaction speed — not only as rendered pages.
- Agent-callable checkout. A checkout endpoint that returns authoritative cart state (items, pricing, taxes, shipping, totals, status) and accepts agent-initiated sessions, ACP-style.
- A payment-authorization answer. A decided path for who authorizes when the agent pays: PSP token support (ACP delegated payments, Stripe Shared Payment Token), AP2-style payment mandates, or network agentic credentials — and the fallback when a mandate fails.
- Consented customer context. Unified profiles with preferences, history, and consent flags, served in real time from the CDP, so agent decisions reflect the whole customer.
- Agent identity and verification. A register of which external agents may transact, verified at the endpoint — the merchant-side half of “know your agent.”
- A dispute path for agent-led orders. Refund and return workflows that accept agent-initiated requests, with an audit trail linking mandate, checkout, payment, and fulfillment.
- Review and content integrity monitoring. Ongoing checks that ratings, reviews, and policy pages match fulfillment reality, because agents act on them as data.
- Agent-channel measurement. Agent-referred traffic, conversion, and margin reported separately from blended totals, with holdouts where agent decisions are live.
Sequencing: items 1, 4, and 8 are the 0–90-day core. Item 2 is the long pole — a multi-quarter engineering effort that items 3 and 6 depend on — so start it first even though it lands last. Items 5 and 7 are ongoing operations, not projects.
FAQ
How is agentic commerce different from traditional e-commerce personalization?
Traditional e-commerce personalization uses recommendation engines to suggest products based on collaborative filtering or content-based matching — “customers who bought X also bought Y.” Agentic commerce deploys autonomous AI agents that manage the entire commerce experience: not just recommendations, but product discovery, pricing, merchandising layout, checkout optimization, and post-purchase engagement. The agents reason about customer intent, make autonomous decisions, execute actions, and learn from outcomes continuously. Traditional personalization is a feature; agentic commerce is an operating model.
What types of retailers benefit most from agentic commerce?
Retailers with large product catalogs (thousands of SKUs), diverse customer segments, and high transaction volumes benefit most because the optimization surface area is too large for humans to manage manually. Fashion, electronics, grocery, and marketplace platforms are natural fits. Specialty retailers with small catalogs and uniform customer bases may find that traditional merchandising and curation are sufficient, as the gains from AI optimization are marginal when the decision space is small.
Does agentic commerce require real-time data, or can it work with batch updates?
Different agentic commerce capabilities have different data freshness requirements. In-session personalization (product discovery, checkout optimization) requires real-time or near-real-time data to respond to current customer behavior. Pricing optimization can work with hourly updates in many cases. Post-purchase engagement can function with daily batch updates. The highest-impact use cases — in-session conversion optimization and real-time product discovery — require the sub-second data access that real-time CDPs provide.
What happens when an AI agent makes a bad purchase?
Impact depends on how reversible the decision was — which is why autonomy should be tiered by blast radius. A misranked search page costs one session; a mispriced product propagates to every channel in minutes. Governed programs contain errors with numeric limits, audit logs of every agent action, rollback for reversible decisions, and human escalation for irreversible ones. Review a log sample weekly: most failures are patterned, and a repeating one signals a missing guardrail, not unusable agents.
What skills does a team need to run agentic commerce?
Merchandising judgment and data discipline matter more than engineering headcount. Someone has to decide what the agents may never do — margin floors, fair-pricing rules, brand-appropriate discounting — and write those constraints in a form the agents can enforce. Beyond that, teams need enough data literacy to interrogate agent telemetry, experiment discipline to separate real lifts from noise, and clear ownership of the escalation path when an agent hits its limits.
Related Terms
- AI Shopping Assistant — The conversational, recommend-and-help-buy application within the broader agentic commerce model
- Conversational Commerce — Selling and serving through chat, messaging, and voice interfaces, the dialogue surface agentic commerce acts on
- Digital Commerce — The channels and infrastructure for online selling that agentic commerce optimizes
- Conversion Rate Optimization — The discipline of improving conversion rates that agentic commerce automates
- Customer Lifetime Value — The metric that agentic commerce agents optimize beyond single-transaction revenue
- Ecommerce Marketing — The marketing strategies for online retail that agentic commerce augments with AI autonomy
- Real-Time CDP — The data infrastructure enabling in-session agentic commerce experiences
This article is also available in: エージェンティックコマースとは？意味と仕組みを解説 · Agentic Commerce: o que é e como funciona