What we learned building SalesBot — HubSpot’s AI-powered chatbot selling assistant

Once I first joined HubSpot’s Conversational Advertising crew, most of our web site chat quantity was dealt with by people. We had a worldwide crew of greater than 100 reside gross sales brokers — Inbound Success Coaches (ISCs) qualifying leads, reserving conferences, and routing conversations to gross sales reps. It labored, nevertheless it didn’t scale.

On daily basis, these ISCs fielded 1000’s of chat messages from guests who wanted product information, had help questions, or have been simply exploring. Whereas we beloved these interactions, they typically pulled focus from high-intent prospects prepared to interact with gross sales.

We knew AI may assist us work smarter, however we didn’t need one other scripted chatbot. We needed one thing that might suppose like a gross sales rep: qualify, information, and promote in real-time.

That’s how SalesBot was born — an AI-powered chat assistant that now handles nearly all of HubSpot’s inbound chat quantity, answering 1000’s of chatter questions, qualifying leads, reserving conferences, and even instantly promoting our Starter-tier merchandise.

Right here’s what we’ve discovered alongside the best way.

How We Constructed SalesBot and What We Realized

1. Begin with deflection. Then, construct for demand.

After we first launched SalesBot, our major objective was to deflect easy-to-answer, low gross sales intent questions (instance: “What’s a CRM” or “How do I add a consumer to my account”). We needed to scale back the noise and unlock people to deal with extra complicated conversations.

We educated the bot on HubSpot’s information base, product catalog, Academy programs, and extra. We are actually deflecting over 80% of chats throughout our web site utilizing AI and self-service choices.

That success in deflection gave us confidence, nevertheless it additionally revealed our subsequent problem. Deflection alone doesn’t develop the enterprise. To actually scale worth, we wanted a software that does greater than resolve — it has to promote.

2. Use scoring conversations to shut the hole.

As soon as we launched deflection, we seen a drop-off in medium-intent leads — those that weren’t able to e book a gathering however nonetheless confirmed shopping for indicators. People are nice at recognizing these moments. Bots aren’t … but.

To shut that hole, we constructed a real-time propensity mannequin that scores chats on a scale of 0–100 primarily based on a mix of CRM knowledge, dialog content material, and AI-predicted intent. When a chat crosses a sure threshold, it’s raised as a certified lead.

That mannequin now helps SalesBot determine high-potential alternatives — even when a buyer doesn’t explicitly ask for a demo. It’s an ideal instance of how AI can floor nuance at scale.

3. Construct to promote, not simply help.

As soon as we’d nailed the foundations of deflection and scoring, we turned our consideration to one thing bolder: turning SalesBot into a real promoting assistant.

We educated it on our qualification framework (GPCT — Objectives, Plans, Challenges, Timeline), enabling the bot to information prospects towards the precise subsequent step: whether or not that’s getting began with free instruments, reserving a gathering with gross sales, or buying a Starter plan instantly in chat.

Now, now we have a software that doesn’t simply reply — it qualifies, builds intent, and pitches like a rep. That shift basically modified how we take into consideration conversational demand era.

4. Select high quality over CSAT.

We shortly realized that conventional chatbot metrics like CSAT (Buyer Satisfaction Rating) weren’t sufficient.

CSAT measures how a buyer feels about their expertise, sometimes by asking whether or not they have been a detractor, passive, or promoter after an interplay. However solely a small portion (lower than 1% of chatters) full the survey. And even when a buyer charges a chat positively, that doesn’t essentially imply the Salesbot was offering a top quality chat expertise.

So we constructed a customized high quality rubric with our top-performing ISCs to outline what “good” truly seems like. The rubric measures components like discovery depth, subsequent steps, tone, and accuracy.

This yr alone, a crew of 13 evaluators manually reviewed greater than 3,000 gross sales conversations. That human QA loop is important. It retains our AI grounded in real-world promoting habits and helps us repeatedly enhance efficiency.

5. Scale globally to spice up efficiencies.

Earlier than AI, staffing reside chat in seven languages was one among our largest operational challenges. It was pricey, inconsistent, and laborious to scale.

Now, we will deal with multilingual conversations world wide, offering a constant expertise regardless of the place somebody’s chatting from. That’s not simply an effectivity win — it’s a buyer expertise improve.

AI has given us true international protection with out overextending our crew, unlocking progress in areas the place headcount merely couldn’t sustain.

6. Construct the precise crew construction.

Success didn’t occur due to one particular person or crew — it occurred as a result of a gaggle of sensible, customer-driven builders got here collectively throughout Conversational Advertising and Advertising Know-how AI Engineering.

Conversational Advertising owned the technique, consumer expertise, and high quality assurance, at all times grounding selections in what would ship the most effective expertise for our prospects. Our AI Engineering companions in Advertising Know-how constructed the fashions, prompts, and infrastructure that made these concepts actual — quick.

Collectively, we fashioned a unified working group with shared targets, a typical backlog, and a rhythm of weekly experimentation. That blend of deep buyer empathy and technical excellence allow us to transfer like a product crew — testing, studying, and enhancing SalesBot with each launch.

7. Strategy automation with a product mindset.

The largest unlock in our journey was embracing a product mindset. SalesBot wasn’t a one-off automation venture. It’s a residing product that evolves with each iteration.

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

These upgrades doubled response pace, improved accuracy, and lifted our certified lead conversion price from 3% to five%.

We didn’t get there in a single day. It took a whole lot of iterations and a tradition that treats AI experimentation as a core a part of the go-to-market movement.

8. People nonetheless matter.

Even with all this progress, some issues nonetheless require a human contact. Right this moment, SalesBot can’t construct customized quotes, deal with complicated objections, or replicate empathy in nuanced conversations — and that’s okay. We’ll at all times be working towards increasing its capabilities, however human oversight will at all times be important to sustaining high quality.

Our brokers and subject material consultants play a core function in our success. They consider outputs, present suggestions, and make sure the system continues to study and enhance. Their judgment defines what “good” seems like and retains our commonplace of high quality excessive because the know-how evolves.

AI’s function is to scale attain and pace — to not exchange human connection. Our ISCs now deal with higher-value applications and edge instances the place their experience actually shines. The objective isn’t fewer people — it’s smarter, extra impactful use of their time.

9. Give your mannequin construction, not simply extra knowledge.

After we first constructed SalesBot, it ran on a easy rules-based system — X motion triggers Y response. It labored for fundamental logic, nevertheless it didn’t sound like a salesman. We needed one thing that felt nearer to an ISC: conversational, assured, and useful.

To get there, we experimented with fine-tuning. We exported 1000’s of chat transcripts and had ISCs annotate them for tone, accuracy, and phrasing. Coaching the mannequin on these examples made it sound extra pure, however accuracy dropped. We discovered the laborious approach that an excessive amount of unstructured human knowledge can truly degrade mannequin efficiency. The mannequin begins remembering the “edges” of what it sees and blurring all the things in between.

So, we pivoted. As a substitute of giving the mannequin extra knowledge, we gave it a higher construction. We moved to a retrieval-augmented era (RAG) setup, grounding the software in real-time context and instructing it when to drag from information sources, instruments, and CRM knowledge.

The result’s a bot that’s considerably extra dependable in complicated gross sales conversations and much better at figuring out intent.

How one can Get Began Constructing an AI Chat Program

In case you’re simply getting began, the largest false impression is that you could bounce straight into AI. In actuality, AI solely succeeds when the muse beneath it’s sturdy. Wanting again at our journey, these three ideas mattered probably the most.

1. Construct the muse earlier than you automate.

AI is just nearly as good because the human program it learns from. Earlier than we automated something, we had years of actual conversations dealt with by expert chat brokers. That reside chat basis gave us:

• Excessive-quality coaching knowledge
• A transparent definition of what “good” seems like
• Patterns to determine what might be automated first

In case you skip this step, your AI gained’t know what “good” is — and it gained’t know when it’s unsuitable.

2. Perceive what your people do nice. Then, educate the AI.

AI can’t replicate the nuances that include human interplay.

Examine your top-performing reps deeply, and ask your self the next questions:

• How do they qualify?
• What indicators do they choose up on?
• What language builds belief?
• How do they recuperate when one thing goes off-script?

Your human crew is your blueprint. All the things nice people do — from tone to timing to discovery — turns into the muse for an AI that may truly promote, not simply reply questions.

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

AI is just not a set-it-and-forget-it venture. Tt’s a product, and the one strategy to scale an AI chat program is to construct a crew that:

• Experiments consistently
• Strikes shortly by way of iterations
• Measures what works (and what doesn’t)
• Treats failures as inputs, not setbacks

An experiment-driven crew turns AI from a one-time launch right into a repeatedly enhancing engine for progress.

The Backside Line

The largest takeaway for me is that this: AI doesn’t exchange nice go-to-market technique — it accelerates it. Your instruments must be a mirrored image of how you use. For us, that’s a mix of know-how, creativity, and buyer empathy to maintain evolving how we promote.