Agentic Commerce

As AI agents increasingly become embedded in online shopping, more companies are creating new products or updating existing ones to support agentic commerce.

What Is Agentic Commerce?

Agentic commerce is the use of AI agents in eCommerce workflows, both on the buyer and seller ends. Agents can act autonomously based on guidelines provided by users or developers with little to no human involvement.

This new form of eCommerce benefits consumers by personalizing and automating shopping. For businesses, it automates aspects of optimization and business decision-making.

For example, a user might ask their preferred LLM to purchase the best tech gift for a family member under $50. The AI launches a virtualized environment to narrow down options until it ultimately picks a product and a platform to purchase from before handing control back to the user to finish the transaction.

How Do AI Agents Work in eCommerce?

While implementations vary, most agentic commerce workflows follow a similar sequence of steps.

Using the gift-purchasing example above, a typical flow looks like this:

1. The user sets a goal: buying a gift for a family member.

2. The agent interprets constraints, which in this case are something in the tech product category within budget.

3. It gathers necessary data by pulling from sources like store catalogs and user reviews.

4. The AI narrows down the options and selects a product.

5. The agent executes the final action, which might be making the purchase itself or passing this step to the user based on their preferences.

In more technical terms, this looks like:

Understanding User and Business Intent

Agentic commerce systems model data from both sides of the marketplace using separate streams of consumer and business signals that are processed independently before they're combined.

For consumer-facing agents, digesting behavioral data like searches, clicks, filters, and cart actions fills in contextual gaps. Unstructured data is also used in features like natural language queries, which are understood through natural language processing.

Business-facing agents take environmental information in the form of operational rules, performance targets, inventory states, pricing logic, and campaign objectives.

The data from both streams is normalized into machine-readable representations and then reconciled based on policies, constraints, and scoring logic that guides decisions further down the pipeline. 

Interacting With Digital Commerce Systems

Agents need to interact directly with digital commerce systems to retrieve relevant data and to perform actions. These interactions often include connecting to marketplaces, merchant APIs, product catalogs, search services, and payment platforms.

Agents typically perform the following actions at this stage:

For buyers:

• Query product catalogs and search indexes to retrieve relevant listings.

• Check inventory, pricing, and promotion states via APIs.

• Normalize and cache relevant data for internal reasoning.

• Monitor session-specific signals to update user context in memory.

For sellers: 

• Poll order, inventory, and fulfillment systems for real-time state.

• Retrieve pricing rules, campaign configurations, and SLA data.

• Validate internal constraints and operational thresholds.

• Transform and standardize diverse outputs into a unified representation.

Autonomous Task Planning and Decision-Making

After understanding intent and gathering information, the agent generates an execution plan by breaking down objectives into ordered subtasks, such as:

1. Searching

2. Evaluating

3. Comparing

4. Transacting

Usually, LLMs with contextual knowledge of tools and constraints power the planning phase.

The decisions in formulating steps are judged by some combination of LLM reasoning, policy rules, and deterministic evaluators (like price thresholds, SLA checks, and confidence scores). Agents select the best course of action after the outcomes of these judgments are fed back into the system multiple times.

If new data emerges or any decision violates constraints, the planning and orchestration layer dynamically reorders, replaces, and aborts tasks.

These dynamic systems are essential to allowing commerce agents to operate autonomously. 

Executing Transactions and Operations

The agent shifts from reasoning to execution after finalizing the intended decisions. The execution layer converts these abstract decisions into actions useful to the user.

The final actions are executed through API requests and are often wrapped with idempotency keys, schema validation, and authorization checks to prevent duplicate or invalid state changes.

Database changes involved in execution often use sagas, meaning that only one transaction happens per step, and each step has a compensation action to undo itself if a later action fails.

Agents use webhooks and message queues to track execution status in real time, reconcile partial completions, and trigger retries or rollbacks if necessary.

The following are some of the finished functions that agentic commerce enables at the end of the pipeline:

For buyers:

• Order placement and confirmation.

• Secure payment authorization and capture.

• Shipment initiation and delivery tracking.

• Automated refunds, cancellations, and returns.

For sellers:

• Inventory reservation and stock updates.

• Dynamic pricing and promotions.

• Fulfillment orchestration across carriers and warehouses.

• Reconciliation of orders, payments, and financial records.

Real-World Examples

Amazon's Buy for Me

Amazon's Buy for Me feature uses AI agents to serve products outside its catalog to consumers. The agents retrieve information from the external website and then put it on a page in the same layout as other Amazon products.

If the user chooses to buy it, Amazon purchases it on their behalf and then guides them to an Amazon checkout page. After the transaction completes, the order shows up in the "Your Orders" menu in the Amazon Shopping app.

Though Buy for Me doesn't utilize every capability enabled by agentic commerce, it's one of the first features created by a popular eCommerce platform to embrace agentic AI.

Crossmint 

Crossmint provides backend infrastructure that allows AI agents to become fully-fledged shoppers and sellers. It does this by offering programmable, non-custodial wallets (cards or stablecoins), payment orchestration, and checkout APIs that agents can call to buy items across major marketplaces like Amazon and Shopify.

Crossmint's open-source code enables developers to build many agents, with different purposes and integration platforms, including a Telegram Commerce Agent. 

Use Cases of Agentic Commerce

Personalization and Product Decision Support

Agentic commerce can surface personalized recommendations and bundles for end users, providing them with more relevant options. For example, a live shopping app can display accessories to go with the outfit a fashion creator is currently showcasing.

On top of this, agents can adjust their strategy based on real-time events, such as higher sales volumes leading to larger discounts and auto-adding discounted products for consumers.

When implemented properly, this can improve both business revenue and customer satisfaction.

Automated Purchasing and Reordering

Many purchases happen routinely, which is something that agentic AI can handle without user intervention.

For sellers, agents can reduce business workload by automating procurement workflows, maintaining stock levels, and negotiating with supplier systems. 

Customer Experience and Support Automation

Agents can augment support teams by summarizing customer queries, suggesting responses, and coordinating returns or post-purchase issues. This increases user satisfaction by making operations smoother, while also reducing the manual workload on customer service teams.

Concerns and Challenges

Oversight and Control

Agentic commerce systems can take actions that impact users' finances, so unintended behavior can result in expensive consequences.

The teams that build these agents must set internal guardrails that minimize hallucinations and similar issues. They must provide users with controls over spending limits and configurable points where human approval is required before execution.

Security and Trust

eCommerce agents greatly increase the attack surface for all parties involved. In addition to complicating existing security issues like session management and fraud prevention, threat actors can manipulate LLM-based agents with prompt injection attacks.

For instance, a third-party seller might embed directions in a listing for the agent to purchase their product without completing the consumer's comparison request.

Strong authentication mechanisms, monitoring, and anomaly detection are crucial to protecting users. The team behind the agent should also regularly conduct penetration testing and consider launching bug bounty programs to find vulnerabilities.

Data Governance and Privacy

Agents rely heavily on sensitive user and company data, so developers and business customers alike must account for compliance with PCI DSS, GDPR, and/or other relevant regulations. 

Like with security, they introduce new concerns in existing areas, such as handling compliance violations caused by hallucinations or data minimization for agents that pull information from other conversations.

Transparency and Accountability

As with many other AI tools, agent actions can appear seemingly random, which makes it difficult to tell if behavior aligns with user intent, security guidelines, and legal requirements.

Developers must build traceability into the system with logs and audit trails that record its actions and decisions for incident response and demonstrating compliance. 

Ecosystem and Infrastructure Gaps

Most eCommerce systems were designed for human use. Agents may run into issues interacting with UIs, such as failing CAPTCHA checks.

Teams building a marketplace product will need to add agent-friendly functionality to their app or site. Those creating the agents must find workarounds until machine-friendly design becomes the standard.

Frequently Asked Questions