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 over a hundred direct sales agents. It worked, but it didn’t measure up.
Every day, these ISCs fielded thousands of chat messages from visitors who needed product information, had support questions, or were just searching. Although we loved these conversations, they were often focused on high-intent prospects ready to engage with sales.
We knew AI could help us do smarter things, but we didn’t want another scripted chatbot. We wanted something think Like a sales rep: Qualify, lead 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 chat questions, qualifying leads, booking meetings, and even selling our starter-tier products directly.
What we’ve learned along the way.
How we built a sales bot and what we learned
1. Start with deviation. Then, create for demand.
When we first launched Salesbot, our primary goal was to answer simple-to-answer, low-sales intent questions (for example: “What is CRM” or “How do I add a customer to my account”) we wanted to reduce the noise and free 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 eliminating over 80% of chats on our website using AI and self-service options.
His success gave us confidence, but it also revealed our next challenge. Business alone does not grow. For value to truly scale, we need a tool that does more than commit — it has to Sell it.
2. Use scoring conversations to close the gap.
Once we introduced corruption, we saw a drop-off in mid-level leads—those who weren’t ready to book a meeting but still showed a buy signal. Humans are very good at seeing these moments. There are no bots yet.
To close this gap, we built a real-time propensity model that scores chats on a scale of 0-100 based on a combination of CRM data, conversation content, and AI-predicted intent. When a chat crosses a certain threshold, it is raised as a qualified lead.
This model now helps the sales bot identify high-potential opportunities—even when a customer doesn’t explicitly ask for a demo. This is a great example of how AI can do that Surface nuance On the scale
3. Build to sell, not just help.
Once we nailed down the basics of defect and scoring, we turned our attention to the argument: turning a sales bot into a true selling assistant.
We trained it on our competency framework (GPCT – Goals, Plans, Challenges, Timeline), enabling the bot to guide prospects to the next step: whether it’s getting started with free tools, booking an appointment with sales, or buying a starter plan directly in chat.
Now, we have a tool that doesn’t just answer — it prepares the qualifiers, intent, and pitches like a rep. This change fundamentally changed how we think about generating conversational demand.
4. Choose criteria higher than CSAT.
We quickly realized that traditional chatbot metrics like CSAT (Customer Satisfaction Score) weren’t enough.
CSAT measures how the customer feels about their experience, usually by asking if they are a detractor, passive, or promoter after the interaction. But only a small fraction (less than 1% of cheaters) complete the survey. And even if a user rates a chat positively, that doesn’t necessarily mean the salesbot is providing a quality chat experience.
So we created a custom quality rubric with our high-performing ISC to define what “good” actually looks like. The rubric measures factors such as depth of discovery, next steps, tone, and accuracy.
This year alone, a team of 13 reviewers manually reviewed more than 3,000 sales conversations. That human QA loop is critical. This keeps our AI grounded in real-world sales behaviors and helps us improve performance.
5. Scale globally to increase efficiency.
Before AI, live chat staff in seven languages was one of our biggest operational challenges. It was expensive, inconsistent and difficult to scale.
Now, we can handle multilingual conversations around the world, providing a consistent experience no matter who is chatting. It’s not just a performance win — it’s a customer experience upgrade.
AI has given us truly global coverage without expanding our team, unlocking growth in regions where headcount simply cannot keep up.
6. Create 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 to engineer conversational marketing and marketing technology.
Exchange Marketing owns strategy, user experience and quality assurance, always underpinning decisions that deliver the best experience for our customers. Our AI engineering partners at Marketing Technology built the models, indicators, and infrastructure that made those ideas real — fast.
Together, we formed a unified working group with shared goals, a shared back blog, and a rhythm of weekly experiences. This blend of deep customer empathy and technical excellence allows us to move like a product team – testing, learning and improving the salesbot with each release.
7. Approach automation with a product mindset.
The biggest unlock of our journey was embracing a product mindset. Salesbot was not a one-time automation project. It is a living product that evolves with each iteration.
Over the past two years, we’ve moved from rule-based bots to a recovery-oriented generation (RAG) system, upgraded our models to GPT-4.1, and added improved qualification and product pitching capabilities.
These upgrades doubled response speed, improved accuracy, and increased our qualified lead conversion rate from 3% to 5%.
We didn’t get there overnight. It took hundreds of iterations and a culture that sees AI experiments as a core part of the go-to-market movement.
8. Humans still matter.
Even with all this progress, some things still require human contact. Today, a salesbot can’t create custom pricing, handle complex objections, or replicate empathy in a meaningful conversation — and that’s okay. We will always work to improve its capabilities, but human oversight will always be necessary to maintain quality.
Our agents and subject matter experts play a fundamental role in our success. They review results, provide feedback, and ensure that the system learns and improves. Their decision defines what “good” looks like and raises our standards of quality as technology evolves.
The role of AI is to scale access and speed — not replace human interaction. Our ISCs now focus on high-value programs and edge cases where their expertise really shines. The goal isn’t less human – it’s a better, more effective use of their time.
9. Give your model structure, not just more data.
When we first built the sales bot, it ran on a simple rules-based system—triggering X-actions. It worked for basic logic, but it didn’t sound like a salesperson. We wanted something that felt close to an ISC: conversational, confident and helpful.
To get there, we experimented with fine-tuning. We extracted thousands of chat transcripts and annotated them for tone, accuracy and phrasing to ISC. Training the model on these examples made it more natural, but reduced in accuracy. We learned the hard way that too much unstructured human data can actually degrade model performance. The model begins to miss the “edges” of what it sees and blurs everything in between.
So, we did. Instead of giving a model more Data, we gave him one better Structure. We moved from a retrieval to collective generation (RAG) setup, grounding the tool in a real-time context and teaching it when to draw from knowledge sources, tools, and CRM data.
The result is a bot that is significantly more reliable in complex sales conversations and far better at identifying intent.
How to start building an AI chat program
If you’re just starting out, the biggest misconception is that you can jump straight into AI. In fact, AI only succeeds when the foundation beneath it is strong. Looking back on our journey, these three principles matter most.
1. Build the foundation before automating.
AI is only as good as the human programs it learns from. Before automating anything, we had years of real conversations through skilled chat agents. This live chat foundation gave us:
- High quality training data
- A clear definition of what “good” looks like
- Samples to identify what can be automated first
If you skip this step, your AI won’t know what’s “good” — and it won’t know when it’s wrong.
2. Understand what your people do. Then, teach the AI.
AI cannot replicate the nuances that come with human interaction.
Study your top performing reps in depth, and ask yourself the following questions:
- How are they qualified?
- What signals do they pick up?
- What language builds trust?
- How will they recover when something is scripted?
Your human team is your blueprint. Everything Great Human does — from tone to discovery — becomes the basis for an AI that can actually sell, not just answer questions.
3. Build an experience-driven, data-driven team.
AI is not a set-it-and-forget-it project. TT is a product, and the only way to scale an AI chat program is to build a team that:
- Experiments constantly
- Moves quickly through repetition
- What Works (and What Doesn’t)
- Treat failures as inputs, not setbacks
An experimentally driven team turns AI from a one-time launch into a constantly improving engine for growth.
The bottom line
The biggest way for me is: AI doesn’t replace great go-to-market strategy—it accelerates it. Your tools should reflect the way you work. For us, it’s a combination of technology, creativity and customer empathy to keep evolving how we sell.
