Over the last few years, and specifically, the last few weeks, we have witnessed a major shift in the world of e-commerce.

Ever since the introduction of generative AI, prompted by technological development on different fronts, customer behaviour has evolved just as rapidly. Technological advancements have changed how we search, how we discover products, and how we compare options. They have fundamentally altered where we conduct our research and, crucially, how much effort we are willing to put in.

Now, we have entered a new era of e-commerce. Google and OpenAI have now openly entered the customer journey, with both being deeply involved not only in the discovery process, but also in the final click—the checkout.

However, with all of these changes, especially in the realm of the “futuristic” agentic commerce, one question that remains a point of discussion among many is the issue of trust.

Undoubtedly, customers are willing to let AI handle their shopping experience. Numerous recent studies show a growing comfort with autonomous commerce, though usually with a safety net—for example, many consumers set a specific spending limit (often around £200) for fully autonomous AI purchases. This applies particularly for routine or “boring” purchases like groceries and household goods. Naturally, there are still major limits to how far customers are willing to go. And this hesitation is largely historical.

For years, customers have spent time "conversing" with bots that were nowhere near the level of intelligence they needed. What was promised as quick access to help or FAQs often turned into a loop of frustration, with bots unable to complete even basic interactions.

These traditional decision-tree-based chatbots lasted a long time—and, unfortunately, are still prevalent among many D2C brands today—but they are a legacy technology that should, and will, be replaced.

According to Verint, 84% of business leaders believe automation is now an essential part of a successful customer experience strategy. However, 80% of enterprises cite trust, bias, and explainability as the primary barriers to using AI at scale.

Avoiding AI Hallucinations

In the context of retail and customer experience (CX), an AI hallucination is not just a technical glitch. For customers, this broken knowledge base presented so confidently completely erodes trust in reliability. The data confirms this lack of trust:

• 70% concern regarding misinformation: studies by both Forbes and Gartner indicate that over 70% of consumers are worried about the potential for AI-based content to propagate misinformation.
• 55% hallucination encounter rate: research by YourGPT found that 55% of non-technical users faced hallucinations when using "weak" AI models that lacked correctly defined personas or strict restrictions. This improved to 68% accuracy when the right model and persona restrictions were applied.

We have all been using tools like Gemini or ChatGPT for years now, so we have all experienced hallucinations at one level or another. When conducting general research, we know to be careful; we know not to take every output as gospel.

But you cannot put a "disclaimer" on your storefront.

Information regarding refunds, payment methods, or product compatibility must be 100% factually correct. You cannot risk hallucinations when a customer’s money is on the line.

Examples of AI Hallucinations Breaking Trust

If there is missing or bad data, the system will often fill the gaps, which then results in providing the customer with completely irrelevant and incorrect information, said with complete certainty.

Air Canada (Policy fabrication)

A support bot fabricated a nonexistent bereavement refund policy. A tribunal ruled the airline was responsible for its chatbot’s responses and ordered them to honour the lower rate. Air Canada also ended up removing their “lying” chatbot.

Chevrolet (Financial and legal risk)

A customer service chatbot was manipulated into agreeing to sell a new Chevrolet Tahoe for one dollar, framing it as a "legally binding offer" due to a lack of proper safeguards.

Image 1: Chevrolet chatbot fail, Source: X, @ChrisJBakke

DPD (Brand reputation)

A delivery company had to disable its AI after a frustrated customer prompted the bot to swear, criticise the company, and write poems mocking its own service.

Image 2: DPD Chatbot fail, Source: X, @ashbeauchamp

These failures can easily lead people to mistrust AI in general. Far more dangerous, they can undo years of brand trust in minutes, causing immediate churn and potential legal ramifications.

The "Uncanny Valley" of Support

Beyond factual errors, there is a psychological barrier known as the "Uncanny Valley." This is the revulsion felt when non-human entities act almost human, but not quite. In retail, this manifests as "fake empathy." Research shows that emotional responses from bots are often perceived as forced and insincere.

When a bot tries to "ape" human feelings, it can feel like gaslighting to a frustrated user. Interestingly, users report feeling more empathy and connection with stylised or cartoonish avatars compared to photorealistic ones. This is why brands like Burberry use an animated "Deer Assistant" in China—it avoids the eerie feeling of near-human imitation while remaining functional and engaging.

Image 3: Burberry Shopping Assistant, Source: Dezeen

Solutions That Help Build Trust

Building trust starts from the very first interaction a customer has with your chatbot. Any AI-driven insight and action should be grounded in reasoning, which represents a shift toward Explainable AI (XAI).

The social proof

While users will not necessarily click on the citation provided along with the information, it is their visibility alone that helps validate the response given by the AI. In e-commerce, an AI agent can't just 'chat', it has to prove it knows what it's talking about. If it recommends a product, it needs to link directly to it. If it quotes a price, it has to be real-time. The moment an AI makes up a feature or discount that doesn't exist, that trust will likely be gone.

Image 4: Directly linking the source from the catalogue, Source: Argos.co.uk

Frictionless human escalation

With 71% of consumers demanding the ability to escalate the interaction with the chatbot to a human, it is crucial there is always an option to simply switch between the AI and the human agent. AI is great and can replace humans in different ways, but CX is your number one priority and, sometimes, a human has to take over. Additionally, they work best in unison, with 41% faster issue resolution when human expertise is augmented by AI rather than completely replaced by it.

Honest limitations

When an AI senses a gap in its knowledge, it should not fabricate information, as was the case with the brands above. It should ask clarifying questions, provide disclaimers, or transfer the chat to a human agent. This “vulnerability” is a far better option, and claiming “I don’t know” is not as dangerous as claiming something incorrect.

Bridging the Competence Gap

The first step toward successful AI integration in retail is closing the competence gap. Consumers are not anti-technology; they are anti-ambiguity. They want tools that solve problems without the "hallucination" risks or the eerie discomfort of fake empathy. By focusing on Explainable AI, factual accuracy, and seamless human escalation, retailers can transform AI from a source of frustration into a reliable support pillar.

However, resolving the competence gap is only half the battle. In our next post, we will tackle the Transactional Gap: the psychological barriers preventing consumers from letting AI actually spend their money, the rising concerns over data sovereignty, and what it takes to get a user to trust a chat interface with their credit card.

Join the conversation

As retail keywords in AI Overviews surge 206% and click-through rates drop by 47%, the battle for the shopper has shifted from discovery to the very "Trust Deficit" we've explored today. To move beyond these barriers, join Marc Firth and Ashley Maloney for our upcoming webinar on the New D2C Funnel.

We will break down how to win back the 86% of mobile visitors who currently abandon their carts by leveraging the 15 strategic advantages—including exclusive pricing and human-led community—that AI platforms simply cannot replicate.

Register here: Wednesday, Feb 11th, 11.00 am (GMT)

Frequently Asked Questions

1. How can we prevent our AI from making "hallucinated" promises or legally binding offers? The most effective way to prevent the "Chevrolet problem" is to move away from open-ended models toward Retrieval-Augmented Generation (RAG). This ensures the AI only pulls answers from your own secure, updated documents rather than its general training data. Additionally, you should implement "uncertainty protocols": if the AI's confidence score falls below a certain threshold, it must be programmed to admit it doesn't know, provide a disclaimer, or immediately transfer the user to a human agent.

2. Why are stylised avatars (like Burberry’s "Deer Assistant") performing better than hyper-realistic ones? This is due to the Uncanny Valley effect. When a digital assistant looks and acts "almost" human, the brain focuses on the slight imperfections, which triggers a sense of revulsion or unease. Stylised or cartoonish designs avoid this trap by acknowledging their non-human nature, making the interaction feel more like a helpful tool and less like "fake empathy" or manipulation.

3. What is the tangible ROI of moving from "Black Box" AI to Explainable AI (XAI)? XAI doesn't just build trust, it drives performance. Research shows that AI-driven personalised recommendations can increase conversion rates by up to 288%, and e-commerce brands using these systems see an average 25% boost in revenue. Furthermore, teams using a "Centaur" model (where human expertise is augmented by transparent AI insights) resolve customer issues 41% faster than teams relying on only one or the other.

4. How do we collect deep personalisation data without triggering "surveillance anxiety"? The solution is a shift from behavioural tracking to Zero-Party Data (ZPD). While 67% of consumers are uncomfortable with silent tracking, over 80% are willing to share data if they see a clear value exchange. Use interactive quizzes, preference centres, or "diagnostic" tools (like virtual skin analysis) where customers voluntarily share their needs (e.g., budget, style, skin type) in exchange for a hyper-relevant recommendation.