Agentic Commerce

Agentic commerce is an approach to buying and selling in which AI agents carry out commercial tasks on behalf of people or businesses, within defined permissions. Instead of manually searching, comparing products, and navigating checkout, a buyer delegates a goal such as “find compatible headphones under $150.” An AI agent—software that selects and executes actions toward a goal—can investigate options, prepare an order, and request approval before purchasing.

How Agentic Commerce Works#

An agentic commerce workflow connects intent, information, and action. A large language model can interpret conversational requirements, while connected tools retrieve product details, check inventory, calculate totals, and submit authorized orders. The model proposes actions; merchant and payment systems validate and execute them.

The agent checks tool responses and adjusts its next step. If a preferred item is unavailable, it might find an alternative or ask whether a different delivery date is acceptable. Stripe’s agentic commerce guide explains the supporting capabilities: product discovery, checkout, payments, and trust. Autonomy varies: some workflows only research products, while others transact under explicit authorization. (stripe.com)

Related Terms and Commerce Protocols#

Agentic commerce describes a commercial application of agentic AI, not a particular model or payment technology. Its boundaries become clearer through several comparisons:

• Recommendation systems: Rank products a customer might like. An agentic system can also check constraints, interact with merchants, and advance a transaction.
• Agentic Commerce Protocol: Defines interfaces connecting buyers, agents, sellers, and payment providers. The protocol supports commerce workflows; it is not synonymous with the broader concept.
• Model Context Protocol: Connects AI applications to tools and external context. It can provide access to commerce services, but does not itself define a complete checkout or payment-authorization workflow. (agenticcommerce.dev)

Conversational commerce describes shopping through dialog; a conversational interface is not necessarily autonomous. Agentic payments are narrower still: they enable authorized payment actions. For example, delegated payment tokens can restrict spending by amount, merchant, and expiration rather than granting unrestricted access to credentials.

Real-World Applications#

Consider two concrete, illustrative workflows.

Personal shopping assistance: A customer requests a dishwasher that fits a specified opening, stays within budget, and arrives before a deadline. The agent compares catalog specifications, checks delivery availability, and presents a final total including installation. It requests approval when dimensions are ambiguous or the total exceeds the authorized limit. The benefit is coordinated decision-making across multiple constraints, not merely generating a product description.

Retail replenishment: A procurement agent combines sales records with computer vision inventory monitoring. A camera-based model flags depleted shelf positions; the agent checks warehouse stock, pending deliveries, and approved suppliers before preparing a replenishment order. Vision provides evidence, while separate procurement logic determines whether purchasing is appropriate. Hidden products or incorrect detections can otherwise cause unnecessary orders.

A Practical Computer Vision Workflow#

Computer vision can supply observations to an agent without becoming the purchasing agent itself. Object detection identifies objects and their locations in an image.

After installing ultralytics with pip install ultralytics, this documented prediction workflow uses Ultralytics YOLO with YOLO26 to produce machine-readable observations:

from ultralytics import YOLO

# Load a pretrained detector for this demonstration.
model = YOLO("yolo26n.pt")

# Analyze a public example image.
results = model("https://ultralytics.com/images/bus.jpg")
result = results[0]

# Serialize detections for a downstream application.
observations = result.to_json()
print(observations)

The JSON output contains detection labels, locations, and confidence scores. This demonstrates the perception-to-software interface—not product identification, stock reconciliation, or checkout. A retail implementation needs task-specific data and custom model training. Ultralytics Platform supports dataset annotation, training, and deployment for that vision component. Commerce authorization remains a separate responsibility. (docs.ultralytics.com)

Risks, Compliance, and Marketplace Preparation#

An agent may misinterpret compatibility, rely on stale prices, or encounter prompt injection—malicious instructions embedded in product pages or other retrieved content. With purchasing permissions, these failures can create financial consequences. Mitigate excessive agency through restricted tools, spending limits, and approval gates enforced outside the model.

Marketplaces should expose accurate catalogs, live availability, explicit return policies, and reliable checkout interfaces. Schema.org Product metadata helps describe products consistently, but does not replace transactional APIs.

Agentic commerce compliance means applying relevant privacy, payment-security, consumer-protection, and recordkeeping obligations to delegated workflows—not obtaining a universal “agentic” certification. Use the NIST AI Risk Management Framework to organize evaluation and oversight. Start with approval-based purchases, preserve transaction records, prevent duplicate orders, and provide clear cancellation and dispute paths. (schema.org)