Agentic Commerce: How AI Agents Are Changing Ecommerce

Ecommerce is moving from answering questions to letting software take action: comparing products, checking inventory, preparing a cart, starting a return, or transferring a support case with its full context. This shift is known as agentic commerce.

The SharkNinja case helps separate practical opportunity from hype. The company connected AI agents to its catalog, orders, warranties, and customer service, and reports that they now handle roughly 20,000 conversations each week. The most useful lesson for an online store is not the platform name. It is the infrastructure SharkNinja had to organize before an agent could answer accurately and act without removing business control.

What is agentic commerce, and how is it different from a chatbot?

A conventional chatbot retrieves an answer or generates text. An agent can reason about an intention, choose a tool, retrieve operational data, and complete a task within defined boundaries. In ecommerce, the difference becomes visible in what happens after the question.

This does not mean every store should enable autonomous purchases. Agentic commerce is a spectrum, from making a catalog understandable to external assistants to offering a first-party agent that completes after-sales tasks. Technical maturity and the risk of each action should determine how far automation goes.

The SharkNinja case: what actually happened

SharkNinja sells Shark and Ninja products across dozens of categories through its own stores, retailers, Amazon, and TikTok Shop. In an interview with CIO Velia Carboni, the company said it rewrote its commerce platform and rolled out Commerce Cloud, Service Cloud, Data 360, Agentforce, Shopper Agent, and order management globally in roughly nine to ten months.

The agent works before and after a purchase. It helps shoppers compare products, recommends accessories, finds replacement parts, checks orders, and registers warranties. A QR code on the packaging opens a setup conversation. That is an especially useful pattern: a direct digital relationship can begin even when the sale happened through a marketplace or third-party retailer.

The numbers, in context

• About 20,000 conversations per week. The diginomica interview describes the general weekly volume; Salesforce refers to weekly volume in the United States.
• 280,000 conversations in the first four months. Salesforce published this figure for the launch of guided-shopping capabilities.
• 11% conversion rate. Salesforce attributes this rate to chats with shopping guidance, but its page does not publish the exact denominator or a controlled comparison.
• 14% year-over-year increase in items added to cart. The figure is associated with the Commerce Cloud launch and should not be attributed to the agent alone.
• 14 countries live. The official June 2026 case study reported this reach, while the later interview describes a broader global rollout.

These are meaningful results, but they mainly come from the solution provider and the company itself. They demonstrate feasibility, not a guarantee that installing a chatbot will reproduce the conversion rate. A sound evaluation needs store-specific metrics: resolution, accuracy, incremental conversion, cost per conversation, escalation, returns, and satisfaction.

The central lesson: data came before AI

SharkNinja did not start with a model. It started with a commerce foundation that could support one. Different product names across regions and brands, duplicate profiles, and siloed systems made it difficult for an agent to form a coherent view. The company unified customer information, catalog data, website activity, and external systems before expanding the agent's actions.

A smaller merchant does not need to copy an enterprise architecture. It does need one reliable source for each fact. If feed prices, product pages, and checkout disagree; if inventory arrives late; or if the return policy only exists in an outdated PDF, the agent will amplify that inconsistency.

AI agents for ecommerce: seven technical requirements

1. A structured, unambiguous catalog

Each product needs stable identifiers, a precise name, variants, attributes, compatibility data, images, price, currency, and availability. Google recommends combining Product and ProductGroup structured data with Merchant Center to improve catalog understanding and verification. Although this guidance addresses Google surfaces, the principle applies to any agent: machine-readable data reduces the need to infer facts from HTML.

For a PrestaShop store, the review should cover combinations, references, EAN/GTIN values, attributes, categories, canonicals, JSON-LD, and consistency between feeds and the visible product page.

2. Near-real-time inventory, pricing, and delivery promises

An agent cannot confidently recommend an out-of-stock product or promise delivery from stale cache data. Inventory needs a source of truth and clear rules when warehouses, ERPs, marketplaces, or physical stores are involved. The system must also distinguish base price, promotions, taxes, shipping charges, and validity dates.

3. APIs with stable contracts

Catalog search, order lookup, cart creation, and return requests should be small, documented, predictable operations. Shopify now publishes UCP capability profiles at /.well-known/ucp and separates tools for catalog, cart, checkout, and orders. Not every store needs UCP today, but the direction is clear: agents perform better through structured interfaces than by simulating clicks on pages built for humans.

An API-first, headless-ready development approach makes that evolution easier even when the storefront remains conventional.

4. Product content that supports decisions

An AI-ready product page does more than repeat keywords. It answers questions about dimensions, materials, installation, consumption, maintenance, compatibility, limitations, warranties, and differences between models. Specific comparison tables and FAQs give agents better evidence than generic descriptions generated at scale.

5. Permissioned access to orders and after-sales data

Checking an order requires customer authentication and a result limited to that account. Registering a warranty or return requires validation of the product, date, policy, and status. An agent should never receive broad database access. It needs least-privilege tools, constrained fields, and audit logs.

6. Human handoff with context

The official SharkNinja case says that after three failed attempts, the system directs a user to a contact page. A real-time transfer carrying full history is described as a future enhancement. The distinction matters: showing a link is not the same as completing a contextual handoff.

A good escalation retains the conversation, authenticated customer, product, attempted actions, and reason for the block. The human team should not make the customer start again. This flow should connect AI with existing customer support and ticketing systems.

7. Security, confirmation, and traceability

OWASP recommends least privilege, explicit authorization for sensitive tools, action previews, autonomy boundaries, and auditable traces. Searching for a product is low risk; cancelling an order, issuing a refund, or charging a payment method is not.

A policy can divide tasks into three levels:

1. Automatic: search products, retrieve public policies, or compare variants.
2. Customer-confirmed: change a cart, save information, or start a return.
3. Human or strongly approved: exceptional refunds, post-payment address changes, manual discounts, or decisions with financial impact.

Store security remains the foundation. Connecting an agent to obsolete modules, shared credentials, or uncontrolled endpoints creates a new attack surface instead of a competitive advantage.

AI for online stores: a realistic roadmap

The first pilot should solve a defined problem, such as finding a compatible spare part, explaining the differences between three ranges, or reducing order-status questions. “Add AI” is not a measurable objective.

Metrics an online store owner should monitor

• Valid autonomous resolution: cases completed without intervention and without later reopening.
• Operational accuracy: correct answers about price, inventory, compatibility, delivery, and policies.
• Incremental conversion: difference against a comparable group or period, not merely orders that touched the chat.
• Order value and margin: a larger cart does not always create more profit.
• Escalation and repetition: how many customers reach a person and how often they must repeat information.
• Action errors: incorrect carts, invalid returns, or unauthorized changes.
• Total cost: licenses, inference, integrations, observability, and review work.

The reported 11% SharkNinja figure is interesting, but each store should define its own funnel: eligible conversations, recommendations shown, clicks, carts, completed purchases, cancellations, and returns. Without this sequence, a conversion number may hide prior selection or incomplete attribution.

Does your ecommerce business need agentic commerce now?

It probably needs to prepare even if it does not yet need an agent that buys. Normalizing the catalog, exposing secure APIs, improving content, unifying inventory, and documenting policies creates immediate value across SEO, feeds, marketplaces, customer service, and automation.

A store is ready for a pilot when it can answer these questions clearly: what is the authoritative source for each fact, which actions may the agent perform, who confirms sensitive actions, how is a conversation transferred, what is logged, and how can an error be reversed? If the answers rely on manual processes or informal knowledge, that is the prerequisite work.

PsDevs can turn this preparation into a technical plan for your current platform: process automation, APIs, ERP or support integrations, catalog optimization, and security controls. Tell us which operation you want to improve, and we will assess whether AI is the right component or whether the foundation should be fixed first.

Sources and further reading

1. Stellagent: analysis of SharkNinja's Agentforce and Shopper Agent rollout.
2. diginomica: interview with SharkNinja CIO Velia Carboni.
3. Salesforce: official SharkNinja agentic shopping and service case study.
4. Google Search Central: product structured data for ecommerce.
5. Shopify Developers: agentic commerce and Universal Commerce Protocol.
6. OWASP: AI Agent Security Cheat Sheet.