Q&A

How does the chatbot qualify leads and route them based on criteria?

The chatbot can walk a visitor through a set of qualification questions, with up to five criteria you define. Those might include things like budget range, company size, geographic region or industry. Each of these questions can be mapped directly to CRM properties in HubSpot, so when the visitor answers, the bot is not just having a conversation – it is actually populating and updating fields on a contact record. That means every qualified visitor becomes a structured new lead your sales team can work with, using the same segmentation, reporting and automation you already have in your CRM.

Once the chatbot has enough information, it evaluates whether the lead is qualified based on the rules you’ve configured. If they meet your criteria, the bot can trigger a routing action such as booking a meeting, updating the lifecycle stage in the CRM or handing them to a specific team. If they are not qualified, you can choose to update their lifecycle stage in a different way or politely close the conversation. The whole flow is designed so that your website stops being just a passive brochure and becomes an actual lead-generation engine with clear routing into your sales process.

How does the knowledge base and self-service experience work in the chat widget?

The chat widget can surface your knowledge base directly inside the conversation, so customers can often help themselves without ever talking to a human. When a knowledge base is connected on a subdomain, the customer can click into a “Help” section and search articles right in the widget. They’ll see trending or popular articles listed, and when they click one it expands in a more readable format inside the chat window, with the option to navigate to additional related articles.

This creates a smooth self-service loop: instead of raising tickets for common questions, visitors can quickly find the answers on their own. Your team is spared from handling repetitive, low-value queries, while the customer gets an immediate, well-structured response. The bot and the knowledge base work together so that general informational questions are answered by content, and the conversational layer is reserved for follow-up, clarification or more complex issues.

What large language model does the customer agent use?

Under the hood, the customer agent is powered by an OpenAI model (ChatGPT) that HubSpot runs within its own environment. In practice, that means you get the conversational ability and reasoning of a modern LLM, but wrapped in HubSpot’s own infrastructure, guardrails and configuration options. From the user’s perspective, you interact with a “HubSpot” bot, but technically it is an OpenAI-based model doing the language understanding and generation.

The important nuance is that HubSpot augments this underlying model with your own context and data rather than letting it roam freely. It doesn’t just search the open web or improvise; it works with the specific content you’ve given it and within the rules you’ve configured. So while the base intelligence comes from an OpenAI LLM, the behaviour is heavily constrained and shaped by your content sources, short answers, CRM data settings and guardrails.

How do we train the customer agent before we go live?

Before deployment, you “train” the agent by feeding it content and explicit Q&A rather than by tweaking neural weights directly. In the customer agent settings, you start by adding content sources: HubSpot-hosted pages, URLs from your website, uploaded files such as product brochures, and other public resources that explain your business. The agent will only use what you provide, so this step effectively defines the pool of knowledge it can draw from when answering questions.

On top of that general content, you can define short answers for specific, common questions – for example opening hours, simple policies or standard explanations. These short answers act like precise overrides: when the customer asks that particular question, the agent can respond with the exact phrasing and details you’ve supplied. Once content and short answers are in place, you can use the built-in testing tool to ask the agent questions and see how it responds before anything goes live, checking that it stays within guardrails and that the tone and accuracy are acceptable.

After deployment, how can we correct wrong answers and improve the agent safely?

Once the agent is live, you don’t retrain the underlying LLM; instead, you continually refine the content and short answers around it. In the customer agent area you can see an overview of conversations and, importantly, a section showing “knowledge gaps” – situations where the bot could not answer a question based on the information it has. You can open those conversations, see exactly what was asked and how the bot struggled, and then create or refine a short answer or add new content so that next time it has a clear, correct response.

If you’re doing formal QA (for example, reviewing 20% of chats in a regulated environment), your process is similar: when you spot a wrong or inadequate answer, you fix the root cause in the knowledge layer rather than “scolding” the model. That might mean adjusting wording in your content, adding a specific question-and-answer pair as a short answer, or tightening instructions so the agent responds differently in similar situations. The bot doesn’t learn automatically from past chats, so each improvement is a deliberate change to its sources and rules, which makes its behaviour more predictable and auditable.

Does any training or usage data from our bot get used to train global models or shared across companies?

The way the system is described, the agent is not self-training on your conversations, and it doesn’t use customer A’s conversations to improve customer B’s bot. The underlying OpenAI model remains what it is; what changes is the context HubSpot feeds into it from your own content, knowledge base and CRM. In other words, the “training” that matters for your agent is done through augmentation with your data, not by updating the core model weights based on your customer interactions.

Practically, that means the bot stays within the guardrails you define. It remembers context within a single ongoing conversation (for example, it recognises you if you are the same logged-in contact coming back), but when a new person starts a chat it’s essentially a blank slate again, using only the content and rules you’ve configured. It doesn’t accumulate behavioural knowledge across tenants, and it doesn’t autonomously evolve by absorbing patterns from other companies’ conversations unless you explicitly mix multiple brands’ content into the same agent, which is something you would control.

Are lead generation and qualification goals handled entirely within the chatbot?

Yes, lead generation and qualification goals are handled inside the chatbot experience itself, but in a conversational way rather than using a traditional form. Agent goals such as “generate leads” or “qualify leads” guide the bot to ask the right questions, capture the key data and then take specific follow-up actions like booking meetings or updating lifecycle stages. To the visitor it feels like a natural conversation, but under the hood each answer is being mapped to CRM properties and evaluated against your qualification logic.

Instead of presenting a static form with fields to fill in, the bot asks for information in a more human sequence and tone. Once it has collected enough details and determined whether the lead meets your criteria, it can route them appropriately: that might mean scheduling time with a sales rep, enriching their CRM record or politely closing things down if they’re not a fit. So the “form” is effectively embedded in the chat, and the action at the end – whether that’s booking, updating or routing – is driven by the agent goals you’ve set.

How do credits work for customer agent conversations across different channels?

Credits are the way HubSpot monetises customer agent functionality, and each conversation consumes a fixed amount. For the customer agent on text-based channels – such as WhatsApp, Facebook Messenger, e-mail and chatbots – one conversation costs 100 credits. The demo chat session shown, for example, used 100 credits as a single conversation, regardless of the number of back-and-forth messages within that window.

There are also time-based rules for when a conversation is considered “the same” versus a new one. On live chat, if you return to the same conversation within 24 hours, it continues under the original 100-credit conversation. For e-mail, that window is 72 hours. WhatsApp has its own rules: there is a 24-hour window after which you must re-engage the user with a template message in line with WhatsApp’s regulations. Other AI tools such as prospecting agents or data agents have their own credit costs, but the key idea is that credits are consumed per conversation per channel, within defined timeframes, rather than per individual message.