Transforming Customer Engagement with AI: The SalesBot Journey

When I first joined HubSpot’s Conversational Marketing team, our website chat was predominantly managed by a dedicated team of human agents known as Inbound Success Coaches (ISCs). With over a hundred sales agents globally, we fielded thousands of chat messages daily from visitors seeking product information, answering support questions, or simply exploring. Although this human touch created valuable interactions, it soon became evident that our approach lacked scalability, particularly as our chat volume continued to grow.

Recognizing an opportunity, we vented our frustrations about scripted chatbots. We wanted an AI solution that could operate like a sales representative—capable of qualifying leads, guiding conversations, and driving sales in real-time. That’s how SalesBot was born, an AI-powered chat assistant that now handles most of HubSpot’s inbound chat volume, fulfilling multiple roles from answering queries to directly selling Starter-tier products.

Here’s a detailed look at the methods and lessons learned from our journey to this innovative AI-driven solution.

1. Start with Deflection, Then Build for Demand

The initial phase of our SalesBot project centered on deflecting low-intent, easy-to-answer questions (think "What’s a CRM?" or "How do I add a user?"). Through comprehensive training of the bot using our knowledge base, product catalogs, and educational resources, we achieved an impressive deflection rate, handling over 80% of chat inquiries through AI and self-service options.

However, while deflection optimized our resources, it also revealed a new challenge: medium-intent leads who exhibited buying signals but weren’t ready to book a meeting. Recognizing that deflection alone wouldn’t grow our business, we understood that we needed to enhance our tool to do more than just resolve queries—it had to sell.

2. Score Conversations to Close the Gap on Demand

To bridge the gap between simple deflection and demand generation, we developed a real-time propensity model. This model scores conversations on a scale of 0-100 based on a combination of CRM data, conversation context, and AI-predicted intent. Once a chat exceeds a predetermined threshold, it’s flagged as a qualified lead.

By employing this scoring system, SalesBot can now identify high-potential opportunities, even when customers don’t explicitly ask for a demo, emphasizing how AI can highlight nuances at scale.

3. Build to Sell, Not Just Support

With deflection and scoring in place, we shifted our focus to making SalesBot a true selling assistant. Leveraging our sales qualification framework (GPCT—Goals, Plans, Challenges, and Timeline), we trained the bot to guide prospects toward appropriate next steps, whether it was starting with free tools, scheduling a meeting, or even purchasing a Starter plan directly through the chat.

This transformative step fundamentally altered our approach to conversational demand generation, as we now have a tool that qualifies, builds intent, and pitches like a seasoned rep.

4. Choose Quality Over Customer Satisfaction Scores

Early on, we discovered that traditional metrics like Customer Satisfaction Scores (CSAT) didn’t provide a comprehensive view of SalesBot’s effectiveness. Although CSAT sheds light on customer feelings post-interaction, it represents feedback from less than 1% of chatters. A positive CSAT rating doesn’t necessarily correlate with quality interactions.

In response, we crafted a custom quality rubric in collaboration with our top ISCs, defining key attributes of a successful interaction such as discovery depth, tone, and accuracy. This year alone, our evaluators manually reviewed over 3,000 sales conversations, lending critical human insights that keep our AI aligned with real-world selling behavior.

5. Scale Globally to Boost Efficiencies

Prior to implementing AI, managing live chat in seven languages posed significant operational challenges—costly and hard to scale, we knew we had to adapt. With SalesBot, we’ve been able to engage in multilingual conversations worldwide, providing a consistent experience for users, regardless of location. This move not only improved efficiency but also enhanced the overall customer experience, enabling growth in regions where staffing constraints had previously hindered our efforts.

6. Build the Right Team Structure

Our success stemmed from the collective effort of various teams—Conversational Marketing handled strategy, user experience, and quality assurance, while our Marketing Technology AI Engineering team developed models and infrastructure. The result? A unified working group with shared objectives, a collaborative backlog, and a rhythm of constant experimentation. Blending deep customer empathy with technical prowess allowed us to progress like a product team, continually improving SalesBot’s capabilities with each iteration.

7. Approach Automation with a Product Mindset

A key transformation in our process was embracing a product mindset for SalesBot. Rather than seeing it as a static automation project, we viewed it as a dynamic product that could evolve. Over the past two years, we upgraded from a rules-based system to a retrieval-augmented generation (RAG) setup, implementing advanced AI with GPT-4.1 for improved qualification and product-pitching abilities.

These advancements doubled our response speed and accuracy while increasing our qualified lead conversion rate from 3% to 5%. Achieving this progress demanded extensive iteration and a culture that prioritizes AI experimentation as integral to our market strategy.

8. Humans Still Matter

Despite our progress, certain elements of the sales process necessitate a human touch. For instance, SalesBot is still unable to craft custom quotes, handle complex objections, or replicate the empathy needed for nuanced conversations. This gap highlights the importance of human oversight for maintaining quality. Our ISCs and subject matter experts are instrumental—they evaluate outputs, provide feedback, and help the AI to learn and improve continually.

AI’s role is to scale our reach and speed, not to replace human connection. Today, ISCs engage with more valuable programs and focus on intricate scenarios that truly leverage their expertise.

9. Give Your Model Structure, Not Just More Data

SalesBot was initially built on a straightforward rules-based system, which fell short of mimicking the natural dialogue of our ISCs. We aimed for a conversational, confident, and helpful tone. Through fine-tuning efforts, we recognized that uncontrolled, unstructured human data could impede model performance.

Instead of feeding the model more data, we provided a clearer structure by pivoting to a retrieval-augmented generation approach. This change drastically improved the bot’s reliability during complex sales conversations and enhanced its ability to discern intent.

Starting Your AI Chat Program

For those considering embarking on a similar journey, it’s crucial to understand that AI implementation thrives on a strong foundation. Reflecting on our experiences, three guiding principles stand out.

1. Build the Foundation Before Automating

A robust AI system is underpinned by high-quality human interactions. We leaned on years of live chat data handled by adept chat agents, which provided:

• High-quality training data
• Clear definitions of desirable outcomes
• Patterns that informed our initial automation efforts

By establishing this foundation, we ensured that our AI understood what constituted a successful interaction.

2. Understand What Your Humans Excel At

AI lacks the nuanced abilities inherent to human interaction. To bridge this gap, we meticulously analyzed our top-performing representatives, questioning:

• How do they qualify leads?
• What signals do they pick up on?
• What language builds trust?
• How do they course-correct in unpredictable situations?

Every effective strategy from our human team served as a blueprint for developing an AI that goes beyond simple responses.

3. Create a Data-Driven, Experiment-Driven Team

AI should not be viewed as a one-off project but as a continuously evolving product. Establish a team culture that thrives on experimentation, agility, and measurement. Success in AI necessitates:

• Constant testing and iterations
• Timely implementation based on data findings
• Viewing failures as learning opportunities

An experiment-driven ethos transforms AI from a static function into a continually improving asset for growth.

As you carve your path toward integrating AI in customer engagement, remember that the goal isn’t merely to replace humans—it’s to enhance their capabilities and drive smarter, more impactful interactions.