Air purifier company Molekule hits 60% resolution rate and lifts CSAT by 42%

Molekule’s support challenges were partly about volume and partly about the nature of the questions. The company’s products generate specific, technical questions that don’t fit neatly into a pre-written FAQ library. A customer wondering whether their filter needs replacement might be asking based on a specific odor, a change in air quality readings, or the age of the unit, and the best answer depends on which.

The existing web chat experience had meaningful limitations:

• Static responses that didn’t adapt to the question: The system matched customer messages to pre-written answers. Multi-part questions got single-part responses. Customers who phrased their question in an unexpected way got irrelevant answers or no answer at all.
• No mobile channel presence: Customers who wanted to reach Molekule from their phones were pushed to web chat, which isn’t the native messaging experience most smartphone users prefer. Apple Messages for Business and Google Business Messages weren’t supported.
• 40% resolution rate that wasn’t improving: The system had no feedback loop. Conversations that the assistant couldn’t resolve went to human agents without any automatic mechanism for capturing what had failed and updating the knowledge base.
• Knowledge base that wasn’t being optimized: The assistant’s accuracy was limited by gaps and imprecision in the underlying knowledge content. There was no systematic way to identify which articles were producing low-confidence or out-of-scope responses.

Quiq built an LLM-powered AI agent grounded entirely in Molekule’s knowledge base, then deployed it first on web chat and later extended it to Apple Messages for Business and Google RCS for Business.

The key elements of the deployment:

• Optimizing the knowledge base with AI: When Molekule’s knowledge base was ingested into the Quiq platform, the system automatically cleaned extraneous content, removed broken links, generated article summaries, and attached likely customer questions to each article. This significantly improved the agent’s ability to retrieve the right information for each query, without requiring the Molekule team to manually rewrite content.
• Retrieval augmented generation (RAG): The assistant uses RAG to ground every response in Molekule’s actual knowledge base, rather than generating answers from general knowledge. This reduces the risk of inaccurate responses and keeps the assistant on-brand and on-topic.
• Multi-part question handling: The LLM retains context across a conversation. When a customer asks two questions in one message, the agent answers both. When a customer follows up with additional detail, the agent incorporates that context into its next response, rather than treating each message as an independent query.
• Disambiguation through follow-up: When a question is ambiguous, the agent clarifies rather than returning a generic answer. A customer who describes a symptom the assistant can’t match to a single product issue gets a follow-up that narrows the diagnosis.
• Feedback loop for knowledge improvement: Quiq built a pipeline with the Quiq AI analyst that flags any responses the AI agent returned with low confidence or out-of-scope classifications. Those flagged responses go back to the Molekule team, who use them to identify knowledge gaps and refine articles. Over time, this loop has produced increasingly precise responses.
• Mobile channel extension: Once web chat performance proved out, Quiq extended the same AI agent to Apple Messages for Business and Google RCS for Business. Customers can now tap a button in Molekule’s mobile app and start an asynchronous messaging conversation using their native messaging app.

Multi-part questions handled in a single response: A customer asks whether their PECO-HEPASilent filter needs to be replaced and whether the unit is still under warranty. The AI agent answers both questions in one response, without requiring the customer to ask again or navigate a menu.

Product identification from description: A customer describes symptoms without naming the exact model. The AI agent evaluates the description against Molekule’s product catalog and identifies the most likely match before proceeding with troubleshooting. It doesn’t require the customer to know the model number.

Graceful escalation with context: When the AI agent reaches the edge of what it can resolve, it doesn’t simply hand off to a human agent and start the conversation over. It passes the full conversation transcript and its own summary of the issue so the human agent can pick up immediately. The customer doesn’t repeat themselves.

The Molekule team’s relationship with the AI agent is now partly editorial. The feedback loop means they receive regular reports on which questions are returning low-confidence responses, and they use those reports to improve knowledge base content. The result is a support system that gets more accurate over time, driven by real customer conversations rather than a periodic content review.

The volume reduction has also reshaped the team’s workload. With 60% of conversations resolved without a human agent, the questions that do reach the support team are meaningfully more complex. The team spends less time on routine troubleshooting and more time on edge cases, returns, and warranty situations that genuinely benefit from human judgment.

“Funneling out-of-scope and low-confidence responses back to our team led to real improvements in knowledge base content and accuracy. The system learns from what it can’t answer.”

– Jay Kershner, VP of Operations

Molekule’s AI agent went from a 40% to a 60% resolution rate, a 50% improvement, while CSAT rose 42%, from 36% to 51%. The agent now handles web chat, Apple Messages for Business, and Google Business Messages from a single unified deployment.

• 60% resolution rate without human agent involvement, up from a baseline of 40%
• 42% increase in CSAT, from 36% to 51%
• Multi-channel coverage: web chat, Apple Messages for Business, and Google Business Messages
• Continuous knowledge base improvement through an automated feedback loop that flags low-confidence responses for team review
• Multi-part questions answered in a single response, with context retained across the full conversation