Introduction

Everyone wants AI on their website. The problem is that few know how to launch it properly, especially when it comes to customer support and repetitive sales questions.

Here’s an inside look at how RevPartners built, trained, and deployed a Customer Agent (named Jarvis) using Breeze inside HubSpot, including how they did it, what worked, what broke, and what they’d do differently.

TL;DR: What will you learn from this guide?

• How RevPartners built and launched a GPT-4-powered Breeze Customer Agent ("Jarvis") in HubSpot
• A framework to go from beta access to a high-performing, fully-trained conversational AI
• What to include in your training data, and what to leave out
• How to handle routing, human handoff, and short answer gaps
• Real early results and the lessons RevPartners learned along the way

What is a Breeze Customer Agent?

Breeze’s Customer Agent is a conversational AI tool baked right into HubSpot. It utilizes large language models (such as GPT-4) to handle repetitive support questions, qualify leads, and route tickets, all without requiring a single line of custom code.

Why did RevPartners build a Breeze Customer Agent in HubSpot?

RevPartners built a Breeze Customer Agent for several reasons, including to handle repetitive customer questions and streamline communication.

What problems did RevPartners have with repetitive customer questions?

RevPartners was fielding the same type of inquiries on a daily basis, such as:

• “What does onboarding include?”
• “Do you support Salesforce migrations?”
• “What is RevOps?”

The questions were simple to answer, but were constant and slowed things down. The sales team was often answering support requests, and support was getting pulled into pre-sales conversations. The result was unwanted friction when attempting to move deals forward.

Why didn’t rule-based chatbots or live chat work for RevPartners?

RevPartners tried rule-based chatbots and live chat. The issues with this approach included chatbots being unable to handle anything outside a script, and live chat requiring someone to be available at all times.

Many teams utilize tools like Drift, Intercom, or Ada to integrate AI into their chat. However, those tools reside outside of HubSpot, which means additional setup, more things to manage, and a higher risk of something breaking. RevPartners wanted something that just worked inside HubSpot, and that’s exactly what Breeze does. Just plug it in and go.

But it wasn’t just about where the AI lived. RevPartners needed an AI that was more capable and could understand real questions, provide accurate answers, and know when to loop in a human.

That’s why the company went with Breeze’s Customer Agent.

Here’s how it stacks up against other support options:

How did Breeze solve real problems without adding complexity?

Breeze’s Customer Agent runs directly inside HubSpot, which meant Jarvis could connect to RevPartners’ existing inbox, help desk, and workflows with no extra tools, syncing issues, or added friction.

But this wasn’t about launching AI just for the sake of it. The goal was to eliminate repetitive tasks from the team’s workload, and it did. Jarvis now handles routine conversations automatically, so RevPartners can focus on onboarding, strategy, and closing real deals.

Every decision in the setup prioritized usability for internal teams and the people asking questions.

As Dharmesh Shah, co-founder and CTO of HubSpot, said: “To really win in the modern age, you must solve for humans.”

That mindset shaped everything from Jarvis’s tone to the fallback logic and routing rules. The goal was to achieve clarity, consistency, and an improved customer experience.

How RevPartners set up its Customer Agent in only a few weeks

Once beta access was granted, RevPartners moved fast. 

In under a month, the team scoped the use case, trained the model, and deployed a live AI agent on the site.

Here's how the company did it:

Step 1: How do you enable the Agent and approve data access in HubSpot?

RevPartners activated the Customer Agent from the Breeze panel inside HubSpot. Then, the team gave Jarvis permission to use conversation and file data so it could generate relevant, informed responses.

Step 2: How do you name the Agent and set its personality?

RevPartners named the Agent Jarvis (yes, like Iron Man’s AI). The role was set to “Sales Rep,” and the tone was set to be witty and helpful, matching the brand’s voice.

Step 3: How do you assign the Agent to a shared inbox and Chatflow?

Jarvis was added to an existing shared inbox and included in a Chatflow. 

The team could now track every message inside HubSpot and make the agent visible to site visitors right away.

Step 4: How do you configure routing and fallback logic for the Agent?

RevPartners created rules to send sales-related questions to the team, trigger tickets for support issues, and display fallback messages when Jarvis was unsure how to answer.

For example, if Jarvis couldn’t confidently respond to a pricing or integration question, it would say something like: “I might need a human to confirm that. Want me to loop in a RevPartners expert?”

From there, it would either open a support ticket or route the chat to the appropriate shared inbox, depending on the context.

All fallback messages were rewritten to stay on-brand and avoid the generic “I’m not sure how to help” responses.

Step 5: How do you customize the Chat Widget and experience?

The company personalized the welcome message and styled the widget to match its site so that the experience felt like RevPartners instead of a third-party tool.

What content should you train the Agent on (and what should you skip)?

When it came time to train Jarvis, the RevPartners team didn’t upload everything they had. They were deliberate about what would help the AI give useful, accurate answers and what would just get in the way.

What content did RevPartners include in the Agent’s training set?

Before uploading anything, the team filtered their content through a simple principle from Dharmesh Shah: “Create value before you try and extract it.”

That meant focusing on assets already used in sales and support conversations built to answer real questions:

• Website pages like RevOps as a Service, HubSpot Migrations, and Pricing
• Knowledge base articles created in HubSpot
• Prewritten answers to common questions that weren’t already covered on the site
• External URLs, like HubSpot’s Knowledge Base, with citations turned on
• Internal files like PDF service overviews

This process gave Jarvis a strong foundation without overloading it with fluff or irrelevant context.

What content did RevPartners intentionally leave out of training?

The team skipped anything that didn’t support a fast, helpful conversation.

• Blogs and webinars meant for learning, not support
• Old or outdated documents
• Anything hard for the AI to read, like disorganized PDFs
• Competitor links or off-brand sources

The goal was to keep the training data focused, consistent, and directly tied to how RevPartners already talks to customers.

Testing, launching, and monitoring Jarvis

Before turning Jarvis loose on the site, RevPartners ran it through real-world scenarios to make sure it wouldn’t embarrass itself … or the brand.

How did RevPartners test the Agent internally?

The team simulated a range of conversations to see how Jarvis would respond. 

RevPartners tested three key categories:

• Website-based questions like “What services do you offer?”
• Short-answer checks like “Do you support statements of work?”
• Random edge cases like “What’s your favorite color?”

Each response was reviewed and categorized as “accurate,” “close but needs improvement,” or “off target.” This helped the team quickly flag gaps in training data and tweak how the agent handled uncertainty.

What did RevPartners change before the Agent went live?

After the initial round of testing, RevPartners conducted a beta test with team members from sales and marketing, with the primary purpose of checking for correct answers, tone, pacing, and fallback logic.

Based on that feedback, the team made adjustments to short answers, refined the fallback copy, and tuned the routing rules to make sure handoffs would go to the right person at the right time.

What metrics did RevPartners monitor after launch?

Once Jarvis went live, RevPartners tracked:

• The number of resolved conversations that didn’t need human follow-up
• The volume and success rate of AI-to-human handoffs
• Visitor feedback (like thumbs up/down ratings)
• Which questions stumped Jarvis and triggered fallback messages
• Response time and how accurately tickets were created from conversations

This ongoing monitoring gave a real-time view into what was working and what needed fine-tuning, and it shaped how the team continued to iterate on both the content and configuration.

What early results did Jarvis produce?

Here’s how RevPartners tracked Jarvis’ success.

What metrics did RevPartners track in the first 1–2 weeks?

Within the first couple of weeks post-launch, the RevPartners team kept a close eye on how Jarvis was handling real conversations. One of the main goals was to determine how many chats could be resolved entirely by the agent without requiring a human to intervene.

The team also tracked the number of unanswered questions. Each of these pointed to a gap in training and usually led to writing or updating a quick reply. These are short, direct responses that help the agent handle high-frequency questions more reliably. Jarvis' ability to escalate questions was also reviewed, especially to confirm that handoffs were routed to the correct person or team.

Feedback tools, such as thumbs-up and thumbs-down buttons, helped gauge visitor satisfaction, while inbox activity and ticket creation provided visibility into whether workflows were working as expected.

What unexpected issues came up?

A few surprises popped up early. For one, some content that had been synced into the training set dropped off without warning and had to be re-synced. It wasn’t clear why, but it reminded the team to review their training sources on a regular basis.

Another unexpected win: short answers turned out to be essential. They helped Jarvis deliver more direct responses and improved the agent's reliability in edge cases.

Oddball user questions also uncovered some tricky routing bugs. These weren’t major failures, but they did expose logic that needed to be tightened to avoid misroutes or fallback dead ends.

What would RevPartners do differently next time?

Even with a successful rollout, RevPartners walked away with a short list of things to change if starting from scratch.

If you're building your own Customer Agent in HubSpot, here's what the team recommends doing differently:

How would RevPartners improve short answer testing?

The biggest lift came from writing and rewriting short answers. Questions like “Do you offer onboarding?” might get asked 10 different ways. If the model only understands one version, you're back to square one. The team would test more variations earlier to ensure the AI captured common synonyms and phrasing patterns before launch.

Why would RevPartners run an internal soft launch next time?

Jarvis performed well under pressure, but it wasn’t flawless out of the gate. A soft launch with sales and marketing would’ve surfaced some of the edge cases sooner, such as confusing fallback responses or unclear handoff triggers. Letting internal users test the agent in a live(ish) environment would’ve tightened things up before external visitors ever saw it.

Why is a final audit of training content important before launch?

Not all training data stays put. During early testing, some pages that had been synced, like pricing and service overviews, were randomly dropped from the training set. It didn’t break everything, but it did cause gaps in answers that had to be patched later. RevPartners now recommends doing a final content audit just before going live.

Why should you build a custom dashboard before launching the Agent?

Out-of-the-box reporting didn’t cut it. Once Jarvis was live, RevPartners needed a way to track conversation volume, ticket routing, and unanswered questions. RevPartners built a custom dashboard using HubSpot data from conversations, help desk messages, and tickets. 

If the RevPartners team did it again, they’d build this dashboard before launch so they could monitor impact in real time, not a week later when things had already gone a little sideways.

Wrapping up: what can startup founders learn from Jarvis?

This RevPartners case study on implementing Breeze’s Customer Agent shows that while you don’t need a dev team or months of prep to launch a conversational AI inside HubSpot, you do need a clear use case, well-scoped content, and a willingness to iterate.

If you follow the steps RevPartners took, from setup to training to post-launch tuning, you’ll be in a strong position to roll out your own Breeze Customer Agent that helps both your team and visitors.