When I first joined HubSpot’s Conversational Marketing team, most of our website chat volume was handled by humans. We had a global team of more than a hundred live sales agents — Inbound Success Coaches (ISCs) qualifying leads, booking meetings, and routing conversations to sales reps. It worked, but it didn’t scale.

Every day, those ISCs fielded thousands of chat messages from visitors who needed product info, had support questions, or were just exploring. While we loved those interactions, they often pulled focus from high-intent prospects ready to engage with sales.

We knew AI could help us work smarter, but we didn’t want another scripted chatbot. We wanted something that could think like a sales rep: qualify, guide, and sell in real-time.

That’s how SalesBot was born — an AI-powered chat assistant that now handles the majority of HubSpot’s inbound chat volume, answering thousands of chatter questions, qualifying leads, booking meetings, and even directly selling our Starter-tier products.

Here’s what we’ve learned along the way.

How We Built SalesBot and What We Learned

1. Start with deflection. Then, build for demand.

When we first launched SalesBot, our primary goal was to deflect easy-to-answer, low sales intent questions (example: “What’s a CRM” or “How do I add a user to my account”). We wanted to reduce the noise and free up humans to focus on more complex conversations.

We trained the bot on HubSpot’s knowledge base, product catalog, Academy courses, and more. We are now deflecting over 80% of chats across our website using AI and self-service options.

That success in deflection gave us confidence, but it also revealed our next challenge. Deflection alone doesn’t grow the business. To truly scale value, we needed a tool that does more than resolve — it has to sell.

2. Score conversations to close the gap on demand.

Once we introduced deflection, we noticed a drop-off in medium-intent leads — the ones that weren’t ready to book a meeting but still showed buying signals. Humans are great at spotting those moments. Bots aren’t … yet.

To close that gap, we built a real-time propensity model that scores chats on a scale of 0–100 based on a blend of CRM data, conversation content, and AI-predicted intent. When a chat crosses a certain threshold, it’s raised as a qualified lead.

That model now helps SalesBot identify high-potential opportunities — even when a customer doesn’t explicitly ask for a demo. It’s a perfect example of how AI can surface nuance at scale.

3. Build to sell, not just support.

Once we’d nailed the foundations of deflection and scoring, we turned our attention to something bolder: turning SalesBot into a true selling assistant.

We trained it on our qualification framework (GPCT — Goals, Plans, Challenges, Timeline), enabling the bot to guide prospects toward the right next step: whether that’s getting started with free tools, booking a meeting with sales, or purchasing a Starter plan directly in chat.

Now, we have a tool that doesn’t just respond — it qualifies, builds intent, and pitches like a rep. That shift fundamentally changed how we think about conversational demand generation.

4. Choose quality over CSAT.

We quickly realized that traditional chatbot metrics like CSAT (Customer Satisfaction Score) weren’t enough.

CSAT measures how a customer feels about their experience, typically by asking whether they were a detractor, passive, or promoter after an interaction. But only a small portion (less than 1% of chatters) complete the survey. And even if a customer rates a chat positively, that doesn’t necessarily mean the Salesbot was providing a quality chat experience.

So we built a custom quality rubric with our top-performing ISCs to define what “good” actually looks like. The rubric measures factors like discovery depth, next steps, tone, and accuracy.

This year alone, a team of 13 evaluators manually reviewed more than 3,000 sales conversations. That human QA loop is critical. It keeps our AI grounded in real-world selling behavior and helps us continuously improve performance.

5. Scale globally to boost efficiencies.

Before AI, staffing live chat in seven languages was one of our biggest operational challenges. It was costly, inconsistent, and hard to scale.

Now, we can handle multilingual conversations around the world, providing a consistent experience no matter where someone’s chatting from. That’s not just an efficiency win — it’s a customer experience upgrade.

AI has given us true global coverage without overextending our team, unlocking growth in regions where headcount simply couldn’t keep up.

6. Build the right team structure.

Success didn’t happen because of one person or team — it happened because a group of smart, customer-driven builders came together across Conversational Marketing and Marketing Technology AI Engineering.

Conversational Marketing owned the strategy, user experience, and quality assurance, always grounding decisions in what would deliver the best experience for our customers. Our AI Engineering partners in Marketing Technology built the models, prompts, and infrastructure that made those ideas real — fast.

Together, we formed a unified working group with shared goals, a common backlog, and a rhythm of weekly experimentation. That mix of deep customer empathy and technical excellence let us move like a product team — testing, learning, and improving SalesBot with every release.

7. Approach automation with a product mindset.

The biggest unlock in our journey was embracing a product mindset. SalesBot wasn’t a one-off automation project. It’s a living product that evolves with every iteration.

Over the past two years, we’ve moved from rule-based bots to a retrieval-augmented generation (RAG) system, upgraded our models to GPT-4.1, and added smarter qualification and product-pitching capabilities.

Those upgrades doubled response speed, improved accuracy, and lifted our qualified lead conversion rate from 3% to 5%.

We didn’t get there overnight. It took hundreds of iterations and a culture that treats AI experimentation as a core part of the go-to-market motion.

8. Humans still matter.

Even with all this progress, some things still require a human touch. Today, SalesBot can’t build custom quotes, handle complex objections, or replicate empathy in nuanced conversations — and that’s okay. We’ll always be working toward expanding its capabilities, but human oversight will always be essential to maintaining quality.

Our agents and subject matter experts play a core role in our success. They evaluate outputs, provide feedback, and ensure the system continues to learn and improve. Their judgment defines what “good” looks like and keeps our standard of quality high as the technology evolves.

AI’s role is to scale reach and speed — not to replace human connection. Our ISCs now focus on higher-value programs and edge cases where their expertise truly shines. The goal isn’t fewer humans — it’s smarter, more impactful use of their time.

9. Give your model structure, not just more data.

When we first built SalesBot, it ran on a simple rules-based system — X action triggers Y response. It worked for basic logic, but it didn’t sound like a salesperson. We wanted something that felt closer to an ISC: conversational, confident, and helpful.

To get there, we experimented with fine-tuning. We exported thousands of chat transcripts and had ISCs annotate them for tone, accuracy, and phrasing. Training the model on these examples made it sound more natural, but accuracy dropped. We learned the hard way that too much unstructured human data can actually degrade model performance. The model starts remembering the “edges” of what it sees and blurring everything in between.

So, we pivoted. Instead of giving the model more data, we gave it a better structure. We moved to a retrieval-augmented generation (RAG) setup, grounding the tool in real-time context and teaching it when to pull from knowledge sources, tools, and CRM data.

The result is a bot that’s significantly more reliable in complex sales conversations and far better at identifying intent.

How to Get Started Building an AI Chat Program

If you’re just getting started, the biggest misconception is that you can jump straight into AI. In reality, AI only succeeds when the foundation beneath it is strong. Looking back at our journey, these three principles mattered the most.

1. Build the foundation before you automate.

AI is only as good as the human program it learns from. Before we automated anything, we had years of real conversations handled by skilled chat agents. That live chat foundation gave us:

• High-quality training data
• A clear definition of what “good” looks like
• Patterns to identify what could be automated first

If you skip this step, your AI won’t know what “good” is — and it won’t know when it’s wrong.

2. Understand what your humans do great. Then, teach the AI.

AI can’t replicate the nuances that come with human interaction.

Study your top-performing reps deeply, and ask yourself the following questions:

• How do they qualify?
• What signals do they pick up on?
• What language builds trust?
• How do they recover when something goes off-script?

Your human team is your blueprint. Everything great humans do — from tone to timing to discovery — becomes the foundation for an AI that can actually sell, not just answer questions.

3. Create an experiment-driven, data-driven team.

AI is not a set-it-and-forget-it project. Tt’s a product, and the only way to scale an AI chat program is to build a team that:

• Experiments constantly
• Moves quickly through iterations
• Measures what works (and what doesn’t)
• Treats failures as inputs, not setbacks

An experiment-driven team turns AI from a one-time launch into a continuously improving engine for growth.

The Bottom Line

The biggest takeaway for me is this: AI doesn’t replace great go-to-market strategy — it accelerates it. Your tools should be a reflection of how you operate. For us, that’s a blend of technology, creativity, and customer empathy to keep evolving how we sell.

Entity-based SEO is a content optimization strategy built around concepts, relationships, and context rather than isolated keyword phrases. Search engines identify entities — distinct concepts, people, places, or things — and connect them through the Knowledge Graph to interpret meaning and determine topical authority.