How to Build AI Agents in HubSpot: A Step-by-Step Guide

From Idea to Live AI Agent in One Week

Two years ago, deploying an AI-powered agent for your business required a development team, months of work, and a six-figure budget. You needed machine learning engineers to train models, backend developers to build integrations, frontend developers to create interfaces, and DevOps engineers to keep everything running. The result was often brittle, expensive to maintain, and limited to large enterprises with deep technical resources.

HubSpot changed this equation with Breeze Agents. Today, a marketing manager with no coding experience can build and deploy an AI agent that generates content, handles customer support, manages social media, or qualifies prospects – all within HubSpot’s interface. The barrier has shifted from “can we build it” to “should we build it, and how do we build it well.” This article provides a step-by-step guide to building each type of Breeze Agent, including the configuration decisions that separate agents that deliver real value from agents that frustrate users and create more work than they eliminate.

Understanding What AI Agents Actually Do

Before building an agent, you need to understand what makes an agent different from a chatbot or an automation workflow. A chatbot follows scripted conversation paths. If a customer asks something outside the script, the chatbot fails. An automation workflow executes predefined actions based on triggers: if X happens, do Y. There is no intelligence, only logic.

An AI agent operates differently. It understands natural language, reasons about context, makes decisions based on available information, and takes actions autonomously. When a customer asks a question the agent has never seen before, it does not fail – it uses its knowledge base and reasoning capabilities to construct a relevant response. When the agent encounters a situation it cannot handle confidently, it escalates to a human with full context rather than providing a bad answer.

Think of the difference this way. A chatbot is a vending machine: push the right button and get the right response. An automation workflow is a factory assembly line: inputs flow through a predetermined process. An AI agent is a knowledgeable employee: it understands the question, uses available resources to find the answer, applies judgment about confidence level, and asks for help when unsure.

Building a Customer Agent: Step by Step

The Customer Agent handles frontline customer support conversations. It uses your knowledge base, help documentation, and historical ticket data to resolve customer inquiries without human intervention. Here is exactly how to build one, based on the process that has worked across dozens of implementations.

Step 1: Prepare Your Knowledge Base (Week 1, Days 1-3)

Your Customer Agent is only as good as the knowledge it can access. Before touching the agent configuration, audit and improve your knowledge base. Start by pulling your top 50 support tickets from the last 90 days. Categorize them by topic. For each category, verify that a knowledge base article exists that fully answers the question. If an article exists but is incomplete, update it. If no article exists, create one.

Rachel, the support leader from our earlier examples, discovered that her knowledge base covered only 60% of the questions customers actually asked. She spent three days with her team creating 35 new articles covering the remaining 40% of common questions. This preparation work was the single most important factor in her Customer Agent’s 42% resolution rate. Without adequate knowledge base coverage, the agent would have escalated most conversations, delivering little value.

Write knowledge base articles with AI consumption in mind. Use clear, structured formatting with descriptive headings. Include step-by-step instructions with numbered lists. Avoid ambiguous language. The AI reads your articles the same way a new support agent would – clear, well-structured content produces better AI responses.

Step 2: Configure the Agent (Week 1, Days 3-4)

In HubSpot, navigate to Automations, then AI Agents, then Customer Agent. The configuration process involves several decisions that determine agent behavior.

Persona and tone. Define how your agent communicates. Should it be formal or conversational? Technical or simple? HubSpot allows you to set tone parameters that match your brand voice. Rachel chose a friendly but professional tone – approachable enough for frustrated customers but competent enough to inspire confidence.

Knowledge sources. Connect your knowledge base, website pages, blog posts, and any other content the agent should reference. Be selective – connecting irrelevant content can cause the agent to provide off-topic responses. Rachel connected her support knowledge base and product documentation but excluded marketing blog posts, which contained promotional language inappropriate for support conversations.

Escalation rules. Define when the agent should hand off to a human. Common escalation triggers include: the customer explicitly requests a human agent, the agent’s confidence in its response drops below a threshold, the conversation involves sensitive topics (billing disputes, account cancellation, data security), or the conversation has exceeded a maximum number of exchanges without resolution. Rachel set her confidence threshold at 80% – if the agent was less than 80% confident in its response, it escalated rather than risking an inaccurate answer.

Channel deployment. Choose where the agent operates: website chat widget, email, or both. Start with chat only. Chat conversations are real-time, which means you can monitor agent performance and intervene quickly if issues arise. Once you are confident in chat performance, expand to email where response quality matters more because corrections are harder.

Step 3: Test Internally (Week 1, Day 5)

Before going live with customers, test the agent extensively with your support team. Have each agent submit 10-15 test questions covering common scenarios, edge cases, and topics you know the agent should escalate. Document every response and categorize it: correct and helpful, correct but could be better, incorrect, or should have escalated but did not.

Rachel’s testing revealed three issues. The agent occasionally referenced competitor features that were mentioned in a knowledge base comparison article – she removed that article from the agent’s knowledge sources. The agent handled multi-part questions poorly, answering only the first question and ignoring the rest – she added guidance in the agent configuration to address all parts of a customer’s question. The agent did not recognize sarcasm or frustration well enough – she lowered the confidence threshold from 80% to 75% for conversations showing negative sentiment indicators.

Step 4: Soft Launch (Week 2, Days 1-3)

Launch the agent to a small percentage of your website visitors – 10-20% of chat sessions. Monitor every conversation in real-time during the first two days. Look for patterns: are there question types the agent handles well? Are there categories where it consistently struggles? Are customers satisfied with the interaction?

Rachel launched at 15% of chat sessions and monitored all conversations for the first 48 hours. She found that the agent excelled at “how to” questions (90% resolution rate) but struggled with troubleshooting questions that required diagnostic follow-up questions (35% resolution rate). She improved the knowledge base articles for troubleshooting scenarios, adding diagnostic decision trees, and the resolution rate for those queries improved to 55% within a week.

Step 5: Full Deployment and Optimization (Week 2, Days 4-5 and Ongoing)

After validating performance at 15%, increase to 100% of chat sessions. Set up ongoing monitoring dashboards tracking: resolution rate (target: 35-50% for first deployment), customer satisfaction for AI-resolved tickets, escalation rate and reasons, and average conversation length. Review these metrics weekly for the first month, then bi-weekly thereafter. The agent improves over time as you refine knowledge base content based on conversation patterns.

Building a Content Agent

The Content Agent generates marketing content: blog posts, landing pages, case studies, and other written assets. Building an effective Content Agent requires different preparation than Customer Agent. Instead of a knowledge base, you need to provide brand context.

Brand voice configuration. Upload examples of your best content – 5-10 blog posts, 3-5 email campaigns, and your brand style guide if you have one. The Content Agent analyzes these examples to learn your brand’s tone, vocabulary, sentence structure, and content approach. The better your examples, the more on-brand the agent’s output will be.

Content briefs. Content Agent works best when given structured briefs: target keyword, audience description, content goal, key points to cover, and desired length. A brief that says “write a blog post about CRM” produces generic content. A brief that says “write a 1,500-word blog post targeting VP-level marketing leaders in B2B SaaS about choosing the right CRM, emphasizing total cost of ownership beyond license price, with a comparison table and a real-world example” produces focused, useful content.

Human review workflow. Never publish Content Agent output without human review. Set up a workflow where AI-generated drafts are automatically assigned to a content editor for review, refinement, and approval. The editor adds original insights, verifies accuracy, adjusts narrative flow, and ensures the content meets quality standards before publication.

Building a Prospecting Agent

The Prospecting Agent identifies, researches, and reaches out to potential customers. Configuration requires defining your ideal customer profile (ICP) precisely.

ICP definition. The more specific your ICP, the better the agent performs. Instead of “B2B companies with 50-500 employees,” define your ICP as “B2B SaaS companies with 50-500 employees in North America, currently using Salesforce or Pipedrive, that have raised Series A or B funding in the last 18 months, with a marketing team of 3-10 people.” This level of specificity gives the agent clear criteria for prospect identification and reduces noise from unqualified targets.

Outreach templates. While the agent personalizes messages, it needs a framework for each outreach sequence. Define the sequence structure (number of touchpoints, timing between steps, channel mix) and provide example messages for each step. The agent uses these as templates while customizing details for each specific prospect based on their company data and context.

CRM integration rules. Configure how the agent interacts with your CRM. Should it create new contact records automatically, or queue them for human review? Should it assign prospects to specific reps based on territory or industry? Should it update existing records or leave them unchanged? These rules prevent the agent from creating data quality issues in your CRM.

Mistakes That Kill Agent Performance

Launching without adequate training data. An agent with a thin knowledge base or vague brand guidelines produces thin, vague output. Invest time in preparation before configuration. The preparation-to-configuration ratio should be approximately 70/30 – spend 70% of your time preparing content and context, and 30% on actual agent setup.

Setting expectations too high initially. A Customer Agent that resolves 35% of inquiries in its first month is performing well. Companies that expect 80% resolution from day one become disappointed and abandon the project prematurely. Set realistic targets: 30-40% resolution in month one, 40-50% by month three, and 50-60% by month six as the knowledge base improves based on conversation patterns.

Not monitoring and iterating. Agents are not set-and-forget. They require ongoing attention: knowledge base updates, escalation rule adjustments, tone refinements, and performance monitoring. Companies that build an agent and stop optimizing see performance plateau or decline. Companies that review agent performance weekly and make incremental improvements see continuous improvement over 6-12 months.

Conclusion

Building AI agents in HubSpot is accessible to teams without technical expertise, but building effective agents requires thoughtful preparation, careful configuration, measured deployment, and ongoing optimization. The process takes about one week from preparation to launch, followed by continuous improvement. Start with the agent type that addresses your biggest operational bottleneck: Customer Agent if support volume is overwhelming your team, Content Agent if content production is your bottleneck, Prospecting Agent if outbound pipeline generation needs acceleration. Each agent you build successfully creates organizational confidence for the next one, eventually creating an AI-assisted operation across marketing, sales, and service that dramatically amplifies your team’s capacity and effectiveness.

Need help building AI agents? Talk to Widelly about expert-led Breeze Agent setup, including knowledge base preparation, configuration optimization, and ongoing performance monitoring.