AI Chatbots: Home Depot and Redfin Face Integration Challenges

Consolidation of AI Chatbots: The Example of Home Depot

Integrating multiple AI chatbots on a website can be problematic, as demonstrated by Home Depot's experience. The company had set up two distinct chatbots: Magic Apron, a product assistant, and Live Chat, a virtual customer service tool. These two tools were accessible in different areas of the site, which caused confusion among users. For instance, a customer looking to check the availability of a product in a local store could only access Live Chat, which was unable to answer their question.

Users should not need to understand the internal architecture of the site to find the help they need. When adding an AI chatbot to your site, ensure that it does not compete with other chatbots. All existing chat options should be consolidated into a single entry point. A chatbot that clearly identifies itself, manages what it can, and escalates questions to humans when it cannot is preferable to multiple bots.

Continuous Accessibility: The Challenge of Redfin

For AI chatbots to be truly useful, they must be accessible throughout the navigation on the site. Redfin faced challenges in this regard, as its chatbot was only available via the search bar. Once a user navigated to a listing or another page, there was no obvious way to return to the chatbot. This limitation frustrated users, who expected to continue their interactions without interruption.

One participant began saving homes that the chatbot showed her. At one point, she navigated away from the chat to review her favorites. When she finished, she wondered how to return to the chat to continue exploring the search results it provided: “I was hoping to spot the little icon [of the chatbot] somewhere. I didn’t see it anywhere to (...) get back in there (...) Usually, I see it in a corner.”

In contrast, the chat from Williams Sonoma followed the user across all pages, allowing her to continue her conversation while browsing options. If your chatbot's goal is to support users through a process that spans multiple pages, ensure it is available on every page. A chatbot that disappears between pages is one that users will stop trying to use.

Clear Communication of Chatbot Capabilities

Another crucial aspect for the effectiveness of AI chatbots is the clear communication of their capabilities. The chatbot from Turo, for example, failed to precisely indicate what it could do, leaving users in the dark. In contrast, the chatbot from Williams Sonoma clearly listed the tasks it could accomplish, providing better guidance to users.

When users open the chatbot for the first time, it should clearly and concisely indicate what it can do. Vague greetings leave users uncertain. The Turo chatbot opens with a generic phrase that essentially says, "Ask me anything," setting expectations that the bot cannot meet — something participants in our study discovered to their detriment when they tried to ask it to find vehicles for a trip.

Indicate the topics or tasks the chatbot can assist with in the introductory message. The Williams Sonoma chatbot, AI Sous Chef, listed finding kitchen utensils, discovering kitchen essentials, and searching for recipes when users first opened it. It did not specify all capabilities but gave users a clear idea of the territory covered.

The opening message should not overwhelm the user with detailed explanations but rather outline what they can expect. The best approach is to tailor the opening message to the page the user is on. For example, when opening Rufus from Amazon from the homepage, it showed broad personalized suggestions (Suggest my next read) and exploratory prompts (What deals would I like?). On a product page, Rufus highlighted questions specific to that product — immediately signaling that it knew what the user was looking at.

Users will not assume that a chatbot knows what they are viewing unless the chatbot explicitly indicates it. Home Depot's Magic Apron can answer specific questions about products, but its interface gives no indication of this capability. One participant assumed that the bot could not see the product he was viewing and manually copied the product names into the chat.

Magic Apron’s capabilities went even further — it could remember products that the user had viewed earlier in the session, not just what they were currently looking at. One participant navigated to a kitchen faucet page and asked the bot if it was compatible with the bathroom sink he had seen earlier. The bot recalled the correct sink and told him they were not compatible — but nothing in the interface suggested it could do this.

Highlighting this capability through suggested questions — such as "Does this faucet match the round porcelain bathroom sink?" — would have shown the user what the bot was capable of and facilitated their shopping experience.

If your chatbot knows what the user is viewing or can remember what they have consulted, inform them of these capabilities. Update the suggested questions or refer to the product they are currently viewing. Users can only benefit from features they are aware of.

Use of Suggested Prompts

Finally, the use of suggested prompts in the form of clickable buttons can greatly enhance interaction with chatbots. Home Depot and Scouting America demonstrated the effectiveness of this approach by offering questions right from the chatbot's opening. However, Scouting America stopped providing these suggestions once the conversation began, leaving users without direction.

There are two places where chatbots can show suggested prompts: when users click on the chatbot for the first time and after the chatbot has provided a response — and when there is more that the user could discover on the topic. Providing suggested prompts reduces the user's effort by removing the burden of formulating questions. To facilitate the selection of a suggested prompt, present them as clickable buttons rather than text. This approach avoids unnecessary typing and allows for multiple suggested prompts without creating a wall of text.

In the opening message, suggested questions can compensate for a more generic greeting. For example, both Home Depot's Magic Apron and Scouting America's Scoutly showed users some possible questions right from the start and allowed them to start the conversation simply by clicking a button.

Unfortunately, Scoutly stopped providing clickable suggested questions once the conversation began, leaving users without guidance on what they could ask next. One participant noted the absence of suggested questions as he started chatting: “[It] doesn’t have the option for follow-up questions, which I think is a very useful thing [...] It could suggest questions or essentially other options to deepen your thinking.”

In contrast, the AI chatbot from Williams Sonoma, AI Sous Chef, continued to offer follow-up questions throughout the conversation. One participant looking for a dough mixer appreciated how the chatbot highlighted considerations she wouldn’t have thought of herself: “Again, I really like their follow-up questions.”