HubSpot has introduced a new way to build automation, and I think it’s one of the more significant developments for businesses already using the platform.
With Agent Hub and Agent Builder, HubSpot users can create AI agents that analyze CRM information, make decisions, and take action. These agents can also work within HubSpot workflows, adding a layer of reasoning to automation that previously depended on predefined rules and conditions.
I recently attended a HubSpot session led by a product manager who demonstrated how to build a custom agent and incorporate it into a workflow.
Having worked with HubSpot since 2012, I’ve built and managed countless workflows for clients. What interests me about this development isn’t simply that AI can automate more tasks. It’s that we’re beginning to see a different approach to what automation can accomplish.
Here’s how it works, what makes it different, and what I’d encourage businesses to consider before getting started.
How Are HubSpot AI Agents Different From Traditional Workflows?
Traditional HubSpot workflows are built around rules.
A contact submits a form, so a workflow creates a task. A deal moves to a new stage, so a notification goes to the deal owner. A customer meets certain criteria, so they’re enrolled in an email sequence.
These workflows can become quite sophisticated, with multiple branches, conditions, delays, and actions. But the logic still needs to be defined in advance.
AI agents introduce the ability to interpret information before determining what to do next.
For example, a traditional workflow can notify a salesperson when a deal is marked closed-lost. It can also require that salesperson to select a loss reason from a dropdown menu.
An AI agent can go further.
It can review the deal’s history, examine available notes and conversations, identify potential reasons the opportunity was lost, and generate an analysis for the sales team.
That analysis can then become part of a workflow.
The distinction is important. Rather than creating a separate workflow branch for every possible scenario, you can introduce an agent that evaluates the available information and produces an appropriate response.
This doesn’t make traditional workflows obsolete. In fact, I think the two approaches will work best together.
Agent Hub vs. Agent Builder: What’s the Difference?
HubSpot introduced two closely related capabilities.
Agent Hub is the central place to discover, manage, and monitor AI agents.
It includes agents built by HubSpot for specific purposes, along with the ability to manage custom agents. Administrators can review activity, identify items requiring attention, and manage the context agents use to perform their work.
Agent Builder is where users create their own agents.
One of the more interesting aspects is that you don’t necessarily need to build an agent through a complicated technical interface.
You can describe what you want to accomplish in plain language using Breeze Assistant. HubSpot helps translate those instructions into an agent with a defined objective, access to relevant information, and specific actions it can perform.
During the session, HubSpot described three essential components of an agent:
- Instructions: What the agent is responsible for accomplishing and how it should approach the task.
- Actions: What the agent is permitted to do, such as retrieving CRM information, generating an analysis, or updating records.
- Knowledge: The information the agent can reference to understand the business and complete its work.
That third component deserves particular attention.
Why CRM Context Matters So Much
An AI agent can only make useful decisions when it has access to relevant, reliable information.
HubSpot’s Context Home helps organize information about a business, including its products, services, customers, processes, and other details that agents may need.
Businesses can also create knowledge vaults containing additional reference materials, such as internal documentation, procedures, and selected CRM records.
This gives agents a more specific understanding of the organization they’re working for.
Consider two companies using the same type of sales analysis agent.
One has well-maintained deal records, consistent sales processes, documented qualification criteria, and detailed conversation histories.
The other has incomplete deal records, inconsistent stage definitions, and very little information about why opportunities were won or lost.
Both companies might be able to create the same agent. But I would expect very different results.
This is where I think the underlying quality of a HubSpot implementation becomes even more important.
I’ve worked with businesses whose CRM data and processes have evolved over many years. Sometimes the biggest challenge isn’t building another automation. It’s understanding how the existing system works, where information is being captured, and whether that information can be trusted.
AI agents don’t eliminate those considerations. They make them more consequential.
A Practical Example: An Agent That Analyzes Lost Deals
The webinar included a demonstration of a custom agent designed to analyze closed-lost deals.
The objective was straightforward: help sales teams understand why opportunities are being lost and identify lessons that could improve future sales conversations.
The process looked something like this:
- A deal moves to the closed-lost stage.
- A HubSpot workflow triggers the custom agent.
- The agent reviews available deal information and applies the company’s sales evaluation methodology.
- It generates an analysis of the loss, including potential contributing factors and recommendations.
- The workflow delivers that analysis to the deal owner.
What I like about this example is that it addresses a problem I’ve seen in many CRM implementations.
Companies collect a tremendous amount of sales information, but much of it never becomes actionable.
A salesperson might document why an opportunity was lost, but unless someone regularly reviews those records, the information often stays buried in the CRM.
An agent could help turn that information into something useful without requiring a manager to manually review every opportunity.
And over time, the same approach could potentially support broader analysis of sales patterns, objections, pricing concerns, or qualification issues.
There is an important caveat, though.
If a salesperson hasn’t documented the deal properly, the agent may have very little to work with. The demonstration addressed this directly: insufficient information can result in a limited analysis.
That’s preferable to presenting unsupported conclusions as facts, but it also reinforces why consistent CRM usage matters.
Building an Agent Is Only Part of the Work
One thing that stood out during the demonstration was how quickly HubSpot could help create a custom agent from a written description.
That’s impressive, but I wouldn’t confuse ease of creation with readiness for production.
In my experience, the most important questions come before the automation is built.
What problem are we trying to solve? What information does the process depend on? What should happen when that information is missing or contradictory? What actions should happen automatically, and which ones require approval?
These questions are particularly important when an agent has permission to modify CRM records or initiate customer-facing actions.
For example, I would approach an agent that analyzes lost deals differently from one that automatically updates deal stages or sends emails to customers.
The first is primarily analytical. The second can directly affect CRM data, sales activity, and customer relationships.
The more authority an agent has, the more carefully I’d want to define its instructions, permissions, and safeguards.
Fortunately, HubSpot has included controls that support this approach.
Testing, Approvals, and Monitoring
During the webinar, HubSpot demonstrated several capabilities designed to help administrators manage agents responsibly.
Users can test custom agents before publishing them and review their outputs. The testing interface also provides visibility into the agent’s activity and an estimate of how many HubSpot credits a production run would consume.
The session explained that these agent tests do not consume credits.
Agent Hub also provides a place to review items that require human attention, including actions awaiting approval.
These controls are important because AI-generated outputs aren’t guaranteed to be correct simply because they sound convincing.
For a new implementation, I’d want to test an agent against a variety of real scenarios, not just one ideal example.
For the lost-deal agent, that might include a well-documented opportunity, an opportunity with very little history, and one with conflicting information.
I’d also want to review whether its recommendations are genuinely useful to the sales team.
An agent that produces a polished summary isn’t necessarily providing business value. The output needs to improve a decision, save meaningful time, or help someone accomplish something they couldn’t easily do before.
What About HubSpot Credits?
Cost is another consideration, particularly as more HubSpot AI capabilities rely on credits.
Custom agents consume HubSpot credits when they perform work in production. The number of credits depends on what the agent does and the information it processes.
During the demonstration, HubSpot showed how users can estimate credit consumption while testing an agent and establish usage limits.
Administrators can also review credit usage across different agents and other credit-consuming features.
This is something I would include in the planning process from the beginning.
If an agent runs whenever a deal closes, how many times might it execute in a typical month? What happens if that volume increases? Is the value of the output worth the ongoing credit consumption?
I wouldn’t recommend deploying agents across an entire CRM simply because the functionality is available.
I’d start with a specific use case, establish a reasonable usage limit, evaluate the results, and expand from there.
When Should You Use an AI Agent Instead of a Workflow?
This is probably the most important practical question for businesses evaluating Agent Builder.
My general approach would be to use a traditional workflow when the process follows clear, predictable rules.
If a contact meets certain criteria and needs to receive a notification, I don’t see a compelling reason to introduce AI.
But when a process requires interpreting unstructured information, evaluating multiple sources, or generating a context-specific recommendation, an agent may offer something that conventional workflow logic cannot easily replicate.
Some examples worth exploring include:
- Analyzing closed-lost deals to identify recurring sales objections.
- Reviewing customer conversations to identify potential service issues.
- Evaluating account activity to suggest appropriate follow-up actions.
- Summarizing information across multiple CRM records to support internal decisions.
- Identifying patterns in customer feedback that might otherwise require manual review.
These are potential applications, not recommendations to automate every step without human review.
The right solution depends on the quality of the available information, the consequences of an incorrect action, and whether the process genuinely benefits from AI reasoning.
My Take: HubSpot Automation Is Entering a New Phase
I’ve spent years helping businesses build automation that reduces manual work, improves consistency, and makes their CRM more useful.
Agent Builder introduces some exciting possibilities, particularly for processes that previously required someone to read through information, interpret it, and decide what should happen next.
But I don’t think the biggest opportunity is to replace every existing workflow with an AI agent.
It’s to identify the places where traditional automation reaches its limits.
There are plenty of processes that should remain simple, predictable, and rule-based. There are others where the ability to analyze information and make context-aware recommendations could provide significant value.
The challenge for businesses will be knowing the difference.
And while HubSpot is making these capabilities increasingly accessible to users without coding experience, successful implementation will still depend on thoughtful process design, reliable CRM data, and appropriate oversight.
That’s the part I’m most interested in exploring as these tools continue to develop.
For businesses already using HubSpot, I’d recommend starting with one well-defined problem, testing what Agent Builder can accomplish, and evaluating the results before introducing agents more broadly.
The technology is changing quickly. The fundamentals of a well-designed CRM still matter.
I’m an independent HubSpot consultant and Solutions Partner. I help businesses design, implement, and improve HubSpot CRM systems, including workflows, automation, reporting, integrations, and ongoing HubSpot administration.