How Scorett Outlet in e-commerce handled 11,395 questions in 9 months through an AI Agent
Scorett one of the biggest retailers in Sweden handled 11,395 questions in 9 months for their Outlet, through their web based AI Agent. The challenge was repetitive customer service and long response times, which we solved through the AI Agent on their e-commerce store, handling 11,395 questions without them having to lift a finger.
The impact
- Handling 11,395 repetitive customer service inquiries.
- Provided responses under minutes for customers.
- Rating of 4.3 out of 5 stars from users.
The problem
Prior to implementing AI into their support workflow, the support team at Scorett only handled customer service via email. Looking to hire more for their support team they instead opted to leverage AI to answer FAQ questions before they go to the human support. This led to them handling on average a thousand customers' questions every month.
They are now equipped with a human-like chatbot which can converse with customers 24/7.
The solution
To help scale their support operations, we designed two primary conversational flows:
- Product Recommendation
- Product Stock Inquiries
In addition, we implemented flows for handling customer complaints and enabling users to easily get in touch with the support team. Below are the key features of the main flows:
Product recommendation system
In the beginning users were able to receive a product recommendation using the “Product Recommendation” button. This helped customers get recommendations, but did not catch the people typing and trying to get a recommendation via sending a text query.
Therefore by using intents we provided the product recommendation option to users that displayed an intent of wanting a product recommendation, but did not choose the product recommendation option.
Through this feature captured more users looking for product recommendations, leading to increased sales.
Product stock system
While the chatbot, powered by an LLM, handled most customer inquiries effectively, it wasn’t connected to real-time inventory systems.
During a period of internal system transitions, we introduced a temporary feature that routed product stock questions to Scorett Outlet’s human support team. This human fallback ensured users could still get timely and accurate responses about product availability—leading to a smoother experience and higher customer satisfaction.
Monthly updates & improvements
Finally, we implemented a range of smaller yet impactful improvements based on insights from monthly reports. By manually reviewing every conversation the chatbot had with users, we gained a comprehensive understanding of customer interactions.
These insights not only informed the major feature updates mentioned earlier but also led to numerous smaller refinements—such as database updates, prompt adjustments, conversational flow tweaks, and API enhancements. Over time, these incremental changes collectively contributed to significant long-term improvements in the performance and reliability of the agent.
Learnings
Model experimentation: going from Claude 3 Haiku → Claude 3.5 Sonnet
We initially powered the chatbot's conversational AI with Claude Haiku, which delivered excellent performance. As shown in the upper-left image, Claude Haiku outperformed GPT-3.5 in our benchmarks, making it the natural choice at the time.
However, in our commitment to continuously improving and delivering the best possible solutions, we transitioned to Claude Sonnet 3.5 shortly after its release. Sonnet 3.5 has proven to outperform not only Haiku but also GPT-4o and even Claude 3 Opus in overall performance.
Integrating an LLM into the chatbot adds flexibility and keeps the experience dynamic rather than static. The chatbot now generates responses based on a blend of internal and public data sources related to Scorett Outlet, allowing it to provide relevant and helpful answers to users.
The result
The bot on ScorettOutlet.se is currently handling ~1000 interactions (answers to questions) every month, whilst having an average rating of 4.3 of 5 stars from users.
Leading to:
- Reduced amount of incoming customer service inquiries.
- Increased customer satisfaction.
- Increased Lifetime Value & Sales.
“At Scorett Outlet we always strive for efficiency, and this chatbot frees up resources for us, allowing us to focus on offering our customers better prices and an amazing customer experience.”—Vidar Tirén, E-commerce manager at Scorett
See it in action
Try it out at scorettoutlet.se (Swedish based system).
Scorett one of the biggest retailers in Sweden handled 11,395 questions in 9 months for their Outlet, through their web based AI Agent. The challenge was repetitive customer service and long response times, which we solved through the AI Agent on their e-commerce store, handling 11,395 questions without them having to lift a finger.
The impact
- Handling 11,395 repetitive customer service inquiries.
- Provided responses under minutes for customers.
- Rating of 4.3 out of 5 stars from users.
The problem
Prior to implementing AI into their support workflow, the support team at Scorett only handled customer service via email. Looking to hire more for their support team they instead opted to leverage AI to answer FAQ questions before they go to the human support. This led to them handling on average a thousand customers' questions every month.
They are now equipped with a human-like chatbot which can converse with customers 24/7.
The solution
To help scale their support operations, we designed two primary conversational flows:
- Product Recommendation
- Product Stock Inquiries
In addition, we implemented flows for handling customer complaints and enabling users to easily get in touch with the support team. Below are the key features of the main flows:
Product recommendation system
In the beginning users were able to receive a product recommendation using the “Product Recommendation” button. This helped customers get recommendations, but did not catch the people typing and trying to get a recommendation via sending a text query.
Therefore by using intents we provided the product recommendation option to users that displayed an intent of wanting a product recommendation, but did not choose the product recommendation option.
Through this feature captured more users looking for product recommendations, leading to increased sales.
Product stock system
While the chatbot, powered by an LLM, handled most customer inquiries effectively, it wasn’t connected to real-time inventory systems.
During a period of internal system transitions, we introduced a temporary feature that routed product stock questions to Scorett Outlet’s human support team. This human fallback ensured users could still get timely and accurate responses about product availability—leading to a smoother experience and higher customer satisfaction.
Monthly updates & improvements
Finally, we implemented a range of smaller yet impactful improvements based on insights from monthly reports. By manually reviewing every conversation the chatbot had with users, we gained a comprehensive understanding of customer interactions.
These insights not only informed the major feature updates mentioned earlier but also led to numerous smaller refinements—such as database updates, prompt adjustments, conversational flow tweaks, and API enhancements. Over time, these incremental changes collectively contributed to significant long-term improvements in the performance and reliability of the agent.
Learnings
Model experimentation: going from Claude 3 Haiku → Claude 3.5 Sonnet
We initially powered the chatbot's conversational AI with Claude Haiku, which delivered excellent performance. As shown in the upper-left image, Claude Haiku outperformed GPT-3.5 in our benchmarks, making it the natural choice at the time.
However, in our commitment to continuously improving and delivering the best possible solutions, we transitioned to Claude Sonnet 3.5 shortly after its release. Sonnet 3.5 has proven to outperform not only Haiku but also GPT-4o and even Claude 3 Opus in overall performance.
Integrating an LLM into the chatbot adds flexibility and keeps the experience dynamic rather than static. The chatbot now generates responses based on a blend of internal and public data sources related to Scorett Outlet, allowing it to provide relevant and helpful answers to users.
The result
The bot on ScorettOutlet.se is currently handling ~1000 interactions (answers to questions) every month, whilst having an average rating of 4.3 of 5 stars from users.
Leading to:
- Reduced amount of incoming customer service inquiries.
- Increased customer satisfaction.
- Increased Lifetime Value & Sales.
“At Scorett Outlet we always strive for efficiency, and this chatbot frees up resources for us, allowing us to focus on offering our customers better prices and an amazing customer experience.”—Vidar Tirén, E-commerce manager at Scorett
See it in action
Try it out at scorettoutlet.se (Swedish based system).