Executive summary

Retail is on the cusp of its next major transformation. Advances in large‐language models (LLMs), multimodal generative AI and seamless integration frameworks have enabled a new paradigm: agentic commerce.

Unlike conventional ecommerce experiences, agentic commerce is driven by AI agents that can reason, act autonomously and collaborate with existing systems to deliver personalised, conversational and context‑aware shopping across the entire value chain.

Early deployments such as Amazon’s Rufus, Shopify’s AI agents and LVMH’s MaIA demonstrate the potential. However, these roll‑outs only scratch the surface. Research shows that demand for AI‑assisted experiences is high: 7 in 10 shoppers already want AI shopping tools. Enterprises are investing heavily: market value for agentic AI in retail was USD 46.74 billion in 2025 and is forecast to reach USD 175.11 billion by 2030. Survey data also shows that 75% of retailers believe AI agents will be vital to compete and 79% of senior executives are already adopting AI agents. 

This paper explains what agentic commerce is, why it is emerging now, where it is gaining traction and how organisations can adopt it responsibly. It provides a maturity framework, technical architecture guidance, market and adoption data, case studies and strategic recommendations, while highlighting risks such as hallucinations, platform lock‑in and regulatory compliance (e.g., the EU AI Ac). Finally, it argues that open, platform‑agnostic frameworks, exemplified by Enthusiast, will give brands the flexibility and control required to thrive in this new era.

1. Defining agentic commerce

Agentic commerce refers to the application of autonomous AI agents across the end‑to‑end retail value chain. These agents do more than chat: they converse in natural language (via text or voice), understand context, perform actions and learn from mistakes.

For example, a shopping agent can help a shopper find the right product, compare features and place an order; a sales agent can quickly understand a customer’s needs, budget, unique requirements through natural conversation, and instantly generate a tailored proposal; a content agent can write and validate product descriptions.

The core characteristics are:‍

▶ Conversational Interaction

• Shoppers interact with the store using natural language (text or voice) instead of traditional search or navigation menus.
• The AI understands context, intent, and follow-up questions.

▶ Autonomous Reasoning & Decision-Making

• Agents can evaluate options, filter results, and recommend the best fit without needing step-by-step user instructions.
• They can weigh multiple factors like budget, style, specs, and current promotions.
  

▶ Multi-System Orchestration

• AI connects to and coordinates across the full commerce stack—product catalog, inventory, CRM, payment gateway, logistics, and more—so it can complete tasks end-to-end.
  

▶ Personalization at Scale

• Uses customer data, purchase history, and preferences to deliver tailored recommendations, offers, and content.
• Can adjust its tone, suggestions, and upsells to match the customer’s profile.
  

▶ Context-Aware Memory

• Remembers previous interactions in the same session (and often across sessions) to maintain continuity and relevance.
• Avoids repetitive questions, keeps track of constraints like “under $200” or “eco-friendly only.”
  

▶ Real-Time Adaptation

• Updates responses based on live data (stock changes, delivery times, trending products).
• Can re-plan or suggest alternatives instantly when circumstances change.
  

▶ Action Execution

• It can help consumers place orders, process returns, schedule deliveries, or apply discounts autonomously once authorized.
  

▶ Continuous Learning & Improvement

• Improves over time by learning from customer interactions, feedback, and outcomes.
  

These capabilities distinguish agentic commerce from earlier chatbots or FAQ search tools, which were limited to scripted responses.

2. Market adoption trends

2.1 Market size

Market analysts are already tracking significant spending on agentic AI for retail. Mordor Intelligence estimates that the “agentic AI in retail and ecommerce” market was USD 46.74 billion in 2025 and projects it will reach USD 175.11 billion by 2030, a compound annual growth rate of 30.2%. The analysis attributes this growth to cost‑to‑serve reduction, automation of supply‑chain operations and improvements in customer experience.

2.2 Enterprise adoption

Survey data reveals that adoption is accelerating but starts to show return on investment (ROI):

• Salesforce’s Connected Shoppers report indicates that 84% of retailers are using AI in some form and 43% are piloting agentic AI. 
• A PwC / Techstrong survey of senior executives finds that 79% are adopting AI agents and 88% plan to increase AI budgets; respondents report benefits like 66% productivity gains, 57% cost savings, 55% faster decision‑making and 54% improved customer experiences.
• Another survey summarised by Amplience and Polestar notes that 43% of retailers are piloting autonomous AI, while 53% are evaluating use cases, and 75% of retailers view AI agents as essential.
• Capgemini reports that the average return on investment (ROI) for generative AI projects is 1.7x.

These data points suggest that adoption is moving from early experimentation to broad deployment, with executives recognising both competitive necessity and tangible ROI.

2.3 Consumer demand and behavioural shift

Consumers are curious about AI shopping, but comfort with AI varies by purchase type and demographic.

Chain Store Age survey of U.S. consumers reveals:

• 69% have used an AI shopping assistant, often via Amazon or ChatGPT, and 74% believe AI can find better deals than they can manually.
• Usage skews toward lower‑risk categories: tech and electronics (47% of AI‑driven purchases), household essentials (28%) and clothing (27%).
• Most users cite time savings as a key benefit (54%) and rely on AI for product information (49%), answering specific questions (43%) and comparing products (40%). At the same time, 61% worry that AI assistants are biased toward certain brands, and one in four users report spending more than planned due to AI recommendations.

Broader multi‑market research shows rising openness to generative AI for shopping while highlighting persistent trust gaps. In Attest’s 2025 survey of consumers across the US, UK, Canada and Australia:

• 47% said they are likely to use tools like ChatGPT or Microsoft Copilot to research purchases, up six points from the previous year.
• Men (52%) and consumers under 50 (around 54%) are more likely to use AI for product research than women (43%) or shoppers over 50 (41%).
• 54% of consumers are likely to engage with an AI chatbot on a brand’s website, but that figure drops to 43% for shoppers aged 50+.
• 43% of consumers trust the information provided by AI tools, and only 33% trust companies to handle the data collected through AI. Younger consumers are more trusting, with 47% of those aged 18–30 trusting AI‑generated information, versus 35% among those aged 50–67.

Nevertheless, expectations for AI‑enhanced shopping are rising. The Coveo commerce report finds:

• 72% of consumers expect their online shopping experience to evolve with generative AI.
• 91% expect online satisfaction to match or exceed in‑store experiences. Many shoppers are willing to share personal data for better deals or improved experiences, yet they also demand transparency and control.

Collectively, these findings paint a nuanced picture: consumer interest in agentic commerce is strong, particularly among younger, tech‑savvy cohorts and for low‑risk purchases, but trust, perceived necessity and privacy concerns remain significant barriers. Retailers that deliver clear utility (price transparency, time savings, tailored recommendations), offer human fallback options and protect customer data will be best positioned to convert curiosity into mainstream behaviour.

3. Macro forces shaping agentic commerce

3.1 Technology readiness

Recent breakthroughs in LLMs and multimodal AI have made it feasible to build agents that can interpret text, images and even voice input.

Models like GPT‑4 and Llama 2 are capable of chain‑of‑thought reasoning and tool usage, enabling them to retrieve information from APIs, databases and search indices.

Techniques such as retrieval‑augmented generation (RAG) and vector search allow agents to ground responses in real data, while orchestration frameworks (e.g., ReAct, LangChain, Enthusiast) support multi‑step planning and error recovery. These innovations reduce the risk of hallucination and unlock more complex tasks than typical chatbots.

3.2 Economic pressure and ROI

Efficiency gains and revenue uplift are key motivators. Bernstein analysts estimate that agent‑driven experiences could lift global ecommerce conversion by 1.5 to 2.5 percentage points per year, unlocking more than US$240 billion in incremental revenue. 

Rierino’s research shows that early adopters of agentic frameworks experience 6–10% increases in revenue and up to 40% improvements in order efficiency. 

Capgemini’s study reports an average ROI of 1.7x across generative AI projects. In addition to front‑end sales, agentic AI reduces operational costs by automating customer support, sales enablement and content production.

3.3 Regulatory and ethical considerations

Regulation is becoming a gating factor for deployment. The EU AI Act introduces a risk‑based framework that classifies AI systems as unacceptable, high‑risk, limited‑risk or minimal risk. 

High‑risk systems must comply with requirements such as risk management, data quality, technical documentation, transparency, human oversight and post‑market monitoring. Prohibited practices include manipulative techniques, exploiting vulnerabilities and social scoring.

Obligations for general‑purpose AI providers (e.g., LLM developers) include providing technical documentation, training data summaries and copyright compliance; providers of models with systemic risk must conduct evaluations and incident reporting.

These obligations take effect on staggered dates between 2025 and 2027, and Member States must establish competent authorities. For brands operating globally, compliance will influence model choice (self‑hosted vs. cloud), data localisation and transparency requirements.

4. Agentic commerce in practice: case studies

4.1 Amazon Rufus

Announced in February 2024, Amazon’s Rufus is a generative‑AI shopping assistant integrated into the Amazon app. 

Rufus answers product questions, explains concepts, compares items and suggests best‑fit options, guiding users from discovery to checkout. The assistant is trained on Amazon’s extensive product catalogue and data from the web. By reducing friction and offering conversational assistance, Rufus exemplifies how agentic commerce can improve conversion and customer satisfaction.

4.2 Walmart Sparky

In June 2025, Walmart introduced Sparky, a generative AI shopping assistant in its mobile app.

Sparky synthesises product reviews, provides occasion‑based recommendations (e.g., back‑to‑school outfits), supports reordering and plans to incorporate multimodal inputs such as photos, audio or video. This multi‑agent strategy positions Walmart for cross‑channel, context‑aware retail, and signals that agentic commerce extends beyond a single chat interface.

4.3 Zalando Assistant

Since October 2024, Zalando’s assistant has operated across all 25 markets and served more than 2 million customers, driving a 40% increase in high‑value interactions. 

The assistant offers personalised fashion advice, curated outfits and interactive discovery, demonstrating how agentic AI can deepen engagement and drive higher‑value purchases.

4.4 LVMH MaIA

Luxury conglomerate LVMH built MaIA, a central AI platform developed with Google that handles more than 2 million monthly queries from 40,000 employees.

MaIA summarises customer interactions, generates personalised messages, optimises prices and manages supply chain decisions. LVMH’s use shows that agentic AI is not limited to consumer interfaces but can accelerate internal operations across product development and luxury service.

4.5 Mastercard Agent Pay

In April 2025, Mastercard launched Agent Pay, a programme enabling AI agents to complete payments securely using tokenisation. Through partnerships with Microsoft, IBM and others, Mastercard verifies AI agents before authorising transactions, ensuring compliance and fraud prevention. This initiative demonstrates that payments infrastructure is adapting to enable agentic transactions, a critical component for widespread adoption.

4.6 Visa and broader payments

Visa also announced support for AI agents (through tokenised transactions) though details are limited due to paywall restrictions. The emergence of Agent Pay and similar initiatives signals that mainstream payments networks recognise the need to accommodate autonomous agents, further legitimising agentic commerce as the next phase of digital payments.

5. Maturity models and frameworks

5.1 Salesforce’s agentic maturity model

Salesforce proposes a four‑level maturity model for agentic commerce:

1. Level 0 – Fixed rules & repetitive tasks: Simple bots or scripts that respond to predefined triggers and perform repetitive actions. Examples include FAQ chatbots or reorder buttons.
   
2. Level 1 – Information retrieval agents: Agents that retrieve information from a single source (e.g., product catalogue, knowledge base) and provide responses but cannot orchestrate complex tasks.
   
3. Level 2 – Simple orchestration: Agents that coordinate multiple steps within one domain, such as filtering products, adding to cart and initiating checkout. They may integrate with one or two tools (e.g., payments API, shipping estimation).
   
4. Level 3 – Complex orchestration across domains: Agents that operate across multiple domains (inventory, customer data, marketing) and can complete end‑to‑end flows, such as upselling related items or rescheduling deliveries.
   
5. Level 4 – Multi‑agent orchestration: A network of agents working together (customer‑facing, support, supply chain) with oversight mechanisms, enabling dynamic collaboration and parallel task execution.
   
   

This model helps organisations benchmark their current capabilities and plan for incremental expansion. It emphasises governance, data quality and performance metrics at each level.

5.2 Adoption readiness stages

Building on industry surveys, we propose the following stages of adoption:

• Exploration: Organisations investigate possibilities and run proofs of concept (POCs). According to Capgemini, 61% of organisations are exploring AI agents.
  
• Piloting: Limited deployments in select workflows. Approximately 23% have pilots underway. Open-source agentic commerce frameworks such as Enthusiast not only are stack-agnostic, but also fully customizable through modular components, allowing companies to pilot with low effort without compromising data.
  
• Partial scale: Integration across several domains, but not yet enterprise‑wide. About 12% of organisations fall here.
  
• Full scale: Agents operate across many processes, with robust governance and monitoring. Only 2% of organisations have reached this level.

These figures illustrate how most companies are still in early stages and that there is a substantial runway for growth.

5.3 Capability map

Agentic commerce spans both customer‑facing and internal capabilities. The following map categorises common use cases:

6. Technical architecture: the agentic commerce stack

Implementing agentic commerce requires a layered architecture:

1. Interface layer: Channels such as chat, voice and multimodal interfaces (e.g., camera uploads) that capture user input and deliver responses.
   
2. Agent layer: The reasoning core that generates plans, invokes tools, monitors progress and handles errors. It may leverage frameworks like ReAct or LangChain and run on LLMs capable of chain‑of‑thought reasoning.
   
3. Tool layer: Connectors to external systems including product catalogues, order management, CRM, CMS and payment APIs. Tools perform actions such as adding to cart, retrieving inventory or updating customer data.
   
4. Validation layer: Mechanisms to ground responses in real data, enforce business rules, check for hallucinations and ensure safety. This includes rule‑based filters and human‑in‑the‑loop review for sensitive actions.
   
5. Hosting layer: Deployment environment (cloud, virtual private cloud or on‑premises) and model serving infrastructure. Organisations may choose between SaaS LLMs or self‑hosted models depending on compliance, latency and cost requirements.
   

Robust logging, monitoring and feedback loops are essential to detect errors and refine agent behaviour over time.

7. Challenges, risks and mitigation strategies

7.1 Accuracy and hallucination control

LLMs can generate plausible but incorrect information. Agents must therefore be grounded in trusted sources (catalogues, knowledge bases) and use retrieval‑augmented generation to reduce hallucination.

Multi‑step reasoning frameworks, tool invocation and explicit error handling help ensure that actions reflect reality. Human oversight should be incorporated for high‑risk or customer‑impacting tasks, such as refunds or price changes.

7.2 Platform lock‑in and vendor dependency

Many agentic solutions are tied to specific platforms (e.g., Shopify). While these solutions can accelerate adoption, they constrain customisation and integrate poorly with custom or composable stacks. An open, platform‑agnostic framework such as Enthusiast allows organisations to avoid vendor lock‑in and maintain control over data and workflow logic. It also enables hybrid deployments combining cloud‑based services with self‑hosted models for compliance or cost reasons.

7.3 Compliance, privacy and ethics

The EU AI Act and other regulations impose obligations on both AI providers and deployers. High‑risk systems must implement risk management, data governance, documentation, human oversight and transparency.

General‑purpose AI providers must provide training data summaries and copyright compliance; providers of models with systemic risk must conduct evaluations and incident reporting. Brands must also ensure that their AI agents respect consumer privacy, avoid discriminatory outcomes and follow ethical guidelines. Failure to comply could result in fines and reputational damage.

7.4 Skills gaps and change management

Deploying agentic commerce requires cross‑functional expertise spanning product management, data engineering, AI/ML, UX design, legal and compliance. Organisations must invest in training, hire specialised talent and foster collaboration across teams. Leadership commitment and a culture of experimentation are critical to move beyond pilots and achieve scale.

7.5 Security and fraud

Agents handling payments or sensitive data are attractive targets for attackers. Payment networks like Mastercard Agent Pay address this by tokenising payment data and verifying AI agents before transactions. Businesses should implement robust authentication, role‑based access control, encryption and continuous monitoring to mitigate fraud and cyber threats.

8. Strategic roadmap and recommendations

8.1 Assess readiness

• Map workflows to identify high‑impact areas where agentic AI can deliver value (e.g., customer discovery, sales support, content generation). Use the capability map in Section 5.3.
  
• Evaluate data quality and integration readiness. Reliable product data, unified customer profiles and accessible APIs are prerequisites.
  
• Align on risk tolerance and compliance requirements (e.g., EU AI Act, consumer privacy laws) to determine whether to deploy cloud or self‑hosted models.
  

8.2 Start with pilots

1. Choose a contained, high‑value use case, such as a conversational product finder or support copilot.
   
2. Implement retrieval‑augmented generation and tool integrations to ensure grounded responses and actionable outcomes.
   
3. Measure success using KPIs, such as conversion rate, average order value, first‑contact resolution time and content throughput.
   
4. Iterate quickly. Adjust prompts, knowledge sources and workflows based on user feedback and performance metrics.
   

8.3 Scale and govern

• Expand across domains after successful pilots, integrating agents into CRM, ERP and supply chain systems. Ensure consistent user experience across channels (web, mobile, voice, social).
  
• Establish governance including approval workflows, monitoring dashboards and incident response plans. Use the maturity model (Section 5.1) as a guide.
  
• Invest in talent and culture. Upskill teams, hire AI architects and embed cross‑functional collaboration.
  

8.4 Choose open, flexible infrastructure

Open‑source frameworks, such as Enthusiast, enable organisations to build agentic commerce capabilities without being locked into a single vendor. Enthusiast provides:

• Integrations for multiple ecommerce platforms (Shopify, Shopware, Medusa, Saleor, Solidus) and APIs.
  
• Model flexibility—compatibility with SaaS LLMs or self‑hosted models like Llama for data control.
  
• Customisable reasoning and validation logic, allowing businesses to align agent behaviour with brand voice and policies.
  

By adopting open technology, companies retain control over data, can comply with regional regulations and avoid dependency on platform roadmaps.

9. Conclusion

Agentic commerce represents a paradigm shift in how products are discovered, purchased and supported.

Market forecasts project high growth, consumer demand is strong and early adopters are realising measurable ROI. However, the transition requires thoughtful planning, robust architecture, cross‑functional collaboration and adherence to emerging regulatory frameworks.

Businesses that embrace open, flexible solutions and invest in the necessary infrastructure and talent will be best positioned to capitalise on this new era. The choice facing brands is clear: rent agentic capabilities from closed ecosystems, or own them via open, composable frameworks that offer control, adaptability and long‑term strategic advantage.

If you’re exploring how to bring agentic commerce into your business, we’d be happy to walk you through real-world scenarios, technical options, and ROI considerations. Schedule a strategy session HERE.

‍

About Enthusiast

Enthusiast is an open-source agentic commerce framework built for commerce teams that need more than an off-the-shelf chatbot. It connects directly to your product catalog, internal knowledge base, and operational tools to power:

• Conversational shopping assistants tailored to your catalog, APIs, and business rules
  
• Support copilots that resolve tickets faster with accurate, grounded responses
  
• Sales tools that surface specs, pricing, and context instantly
  
• Automated content creation and validation at scale
  
  

Built by Upside, a team with over 8 years of hands-on e-commerce experience, Enthusiast combines technical depth with industry insight. It’s designed for both cloud-hosted and self-hosted deployments, giving you full control over your data, model choice, and integration roadmap. The framework comes with pre-built connectors for Shopify, Shopware, Medusa, Solidus, Saleor, and more, while remaining fully customizable to your workflows.

Learn more: upsidelab.io/tools/enthusiast | GitHub repository