Stop Building Chatbots: Why Multi-Agent AI Architecture is the Future of Retail

In February 2024, Klarna announced a groundbreaking milestone: its AI assistant had handled 2.3 million customer conversations in a single month. The assistant had effectively replaced the work of 700 full-time agents, reduced resolution times from 11 minutes to under two, and was projected to save the company $40 million annually. The announcement, co-signed by OpenAI, was celebrated as a triumph of AI-driven support automation and conversational AI architecture in customer service.

Fast forward 15 months, and the narrative had shifted. Klarna’s CEO, Sebastian Siemiatkowski, admitted to Bloomberg that the company was rehiring human agents. The AI-only approach, he confessed, had resulted in “lower quality” support. By Q1 2025, Klarna’s losses had quadrupled compared to the previous year.

So, what went wrong? Did the technology fail? Not exactly. Klarna’s chatbot was competent at handling simple, bounded tasks, such as order status inquiries and payment schedules. But when it came to complex issues—billing disputes, sensitive payment situations, or conversations requiring judgment—the chatbot fell short. It was adequate at everything but excellent at nothing, which ultimately translated to mediocrity.

This failure highlights a critical flaw in the current approach to conversational AI: a single chatbot trying to do it all. The problem isn’t the AI model itself but the conversational AI architecture behind it. Retailers need to stop building monolithic chatbots and start embracing multi-agent AI architecture powered by autonomous AI agents.

The One-Employee Store: A Flawed Analogy for Chatbots

Imagine walking into a well-run retail store. You’ll find a greeter at the door, a product specialist who knows the merchandise inside out, a cashier, and perhaps a returns-desk associate. In high-end retail stores like Sephora or Nordstrom, you’ll encounter even more specialized roles, such as Beauty Advisors with expertise in skincare, fragrance, or hair care. This specialization isn’t accidental—it’s strategic. Specialists convert better, with retail sales specialists earning a 42% premium over generalist associates. This highlights the core debate of chatbot vs AI agent approaches.

Now, imagine replacing all these roles with one person. It’s absurd, right? Each role requires unique skills, knowledge, and conversational styles. A greeter optimizes for warmth and speed, while a product specialist focuses on understanding customer needs and curating options. Yet, this is exactly what most retailers do when they deploy a single chatbot to handle everything from “Where’s my package?” to “Help me find a dress for my sister’s wedding.”

The result? Chatbots tuned for quick, transactional queries feel cold and unhelpful during exploratory conversations. Conversely, chatbots designed for empathetic, nuanced interactions are too slow for simple tasks. This one-size-fits-all approach is not just inefficient—it’s a recipe for failure.

Why Better Models Won’t Solve the Problem

It’s tempting to think that the next generation of AI models—GPT-5, Claude OPUS, or Amazon’s Grok—will solve these issues. But research suggests otherwise. A May 2025 study titled “LLMs Get Lost in Multi-Turn Conversation” found that performance degrades by 39% in extended, underspecified conversations—exactly the kind retail chatbots handle daily. Even with larger context windows or advanced prompting techniques, the degradation persists.

IBM’s LongFuncEval study further quantified this: as the number of tools available to a single agent increases, accuracy drops by 7% to 85%. Similarly, as conversations grow longer, retrieval accuracy declines by 13% to 40%. These aren’t limitations of the models themselves—they’re limitations of the single-agent architecture. A smarter model in the same architecture will hit the same ceiling, just faster.

What Leading Retailers Are Building Instead

The largest retailers have already moved beyond the single-chatbot paradigm. They’re adopting a multi-agent AI retail model that mirrors the specialization found in physical stores.

Take Walmart, for example. In June 2025, they launched Sparky, an agentic assistant that coordinates specialized sub-agents for catalog search, pricing, fulfillment status, and product reviews. Similarly, Amazon’s Rufus, Lowe’s Mylowe, and Ikea’s generative AI design tool all follow a similar pattern: specialized agents, each handling a specific task, orchestrated by a central layer.

Here’s how it works:

1. Orchestrator Agent: At the top of the system, an orchestrator agent interprets customer intent, routes queries to the appropriate sub-agent, and manages transitions.
2. Specialized Sub-Agents: Each sub-agent is designed for a specific task:
   • A discovery agent helps customers find products based on their needs and preferences.
   • A comparison agent evaluates product attributes and trade-offs.
   • A transaction agent handles cart management, checkout, and payment.
   • A post-purchase agent manages tracking, returns, and re-engagement.

The key innovation lies in how context flows between these agents. Instead of forwarding entire conversation histories, the orchestrator passes only the relevant information—like a structured chart in a hospital. This reduces token usage, minimizes errors, and ensures seamless handoffs. The result? A system that feels like interacting with a knowledgeable friend rather than navigating a frustrating menu tree.

The Organizational Challenge of Multi-Agent AI

Adopting multi-agent AI architecture isn’t just technically challenging—it’s organizationally demanding. Most chatbots today are owned by a single department, such as customer care or digital marketing, and are measured by clear KPIs, such as deflection rate and containment. Multi-agent systems, however, require cross-departmental collaboration.

For example:

• The discovery agent needs access to merchandising and catalog data.
• The recommendations agent relies on marketing’s customer segmentation.
• The order status agent requires integration with supply chain and fulfillment systems.
• The returns agent must align with operations and finance.

This level of integration demands shared budgets, joint KPIs, and collaboration across traditionally siloed teams. A 2025 Deloitte survey found that AI-driven outcomes peak when decision rights are distributed across complementary leaders rather than concentrated in a single function. However, this requires a cultural shift that many organizations find uncomfortable.

Three Objections to Multi-Agent AI (and Why They’re Wrong)

1. “Our chatbot works fine.” If your chatbot handles a defined set of tasks well, that’s great. But is it quietly becoming the default solution for every customer problem? If so, it’s only a matter of time before it hits its ceiling.
2. “Customers don’t want AI.” True, many customers prefer human interaction. A Five9 survey found that 75% of consumers still prefer a human for customer service. However, multi-agent systems create clear escalation points, allowing AI to handle routine tasks while seamlessly offloading complex issues to human agents.
3. “It’s too complex.” Multi-agent systems are indeed more complex to build and manage. But starting small—say, with two agents handling distinct tasks—can help organizations build confidence and refine their orchestration layer before scaling up.

Where This Is Heading

The industry is shifting from static bots to Agentic AI in retail, where AI agents for retail collaborate dynamically through an AI orchestration layer. This shift is not incremental—it’s transformational. Multiple studies place agentic AI at the peak of innovation, while NRF 2026 has made “agentic commerce” its central theme.

At Hexaware, we’ve been building toward this future.  Our Personal Shopping Assistant is a multi-agent AI architecture system delivering deep AI personalization in retail. It’s a multi-agent system designed to deliver seamless, personalized shopping experiences. Each agent in the system has a specific role, from intent recognition to product retrieval, all coordinated by an orchestration layer that ensures the customer never repeats themselves.

For example, a shopper looking for an interview outfit might say, “I need something business casual for a tech panel in San Francisco.” The system knows her size, preferred brands, and past purchases. The agents take a guided, step-by-step approach, focusing on one category at a time and asking targeted questions to recommend the best products per category. One single conversation to curate a full outfit—blazer, top, trousers, shoes, and bag—while checking for fit, color coherence, and shipping timelines.

One conversation. No search results page. No “Did you mean?”

The Future of Retail AI

If your current chatbot feels like it’s not quite working, it probably isn’t. And the solution isn’t a better bot—it’s a better architecture. At Hexaware, the future belongs to multi-agent AI architecture, where AI agents for retail collaborate, specialize, and deliver truly intelligent experiences.