HubSpot's sales representatives had good reasons to build their own AI prospecting agents.
It was fun. And plenty of them believed they had found a better way to research prospects and decide who to contact.
So HubSpot let people experiment with AI. Then its internal results pointed in a different direction: Jon Dick, HubSpot's Chief Customer Officer, says prospecting systems built across the company and tuned with shared context and evaluations consistently outperformed the versions individual representatives built for themselves.
That comparison captures a central lesson from HubSpot's larger AI transformation. To decide where agents belonged, the customer-platform company mapped the work across attracting prospects, engaging buyers, and delighting customers. Then it chose the largest constraints, assigned AI systems bounded jobs, and judged them against go-to-market outcomes.
Dick leads HubSpot's global sales and customer success organizations. In Episode 234 of The Artificial Intelligence Show, as part of our special AI Transformations series presented by Google Cloud, he explained to us the decisions behind that system and why a useful agent needs more than a capable model.
When HubSpot moved beyond isolated use cases, it mapped the customer journey and the work its teams performed at each step.
This mattered because each part of the journey had a different constraint. Marketing needed to reach buyers who were shifting from traditional search to AI answer engines. Sales representatives lost time to account research, administration, and coordination. Customer success teams struggled to give individualized help across hundreds of thousands of customers.
HubSpot could not rebuild every step at once.
"Pick something to focus on," Dick advises. "Pick a real problem for your business and focus on trying to solve that problem versus doing everything at once."
For a sales leader who lacks enough people to build pipeline, that creates a concrete choice. The company can hire more business development representatives, or it can assign a prospecting agent the repeated research and signal-monitoring work that consumes their time. The job comes first. The agent is one possible way to change it.
HubSpot's current systems do not ask one agent to run the entire customer journey. They divide the work.
In sales, a prospecting agent handles account research and monitors signals. Once a deal is underway, Guided Success, a custom assistant built in HubSpot's Breeze assistant, gives the representative HubSpot's own guidance on how to win deals.
Customer success managers use an assistant that can trigger agents to complete parts of the account work. Dick describes the goal in practical terms: Clear more work from the manager's plate so that person can spend more time with customers.
The division of work is specific. The prospecting agent handles repeated research. The sales representative works the deal with guidance from the assistant. The customer success manager spends time with customers while the assistant coordinates supporting tasks.
The prospecting experiments gave HubSpot a direct comparison between two ways of building.
Individual representatives created assistants and agents around their own ideas. Dick says HubSpot consistently saw globally built systems outperform those individual versions.
"The stuff that has been globally built and really tuned with the right context and the right evals and everything just really outperforms the stuff that individual reps build," Dick says.
More broadly, Dick says HubSpot's products bring two forms of context to customer-facing agents: data and workflows already stored in a customer's account, and HubSpot's accumulated experience across marketing, sales, and service.
The internal prospecting agent is a separate example of the value of a shared build. Dick says it booked about 10,000 meetings in the previous quarter.
A company does not need to force every step of the journey through one model.
"We look for the best model for the best job," Dick says, of the company’s go-to-market teams.
Separately, HubSpot's product can orchestrate work across different models. Today, that product capability can help pursue the best result. Over time, Dick says, it may also help manage costs and other concerns.
This puts model selection in its proper place. A team first defines the business problem, the work to be done, and the result that matters. Then it can evaluate which model belongs inside that system.
A well-defined job and strong context still do not settle what good performance means.
HubSpot learned that in customer support. Its teams initially focused on resolution rate, the share of customer issues the AI support experience resolved. But customer satisfaction moved in the wrong direction. HubSpot changed what it optimized and tuned the experience around customer satisfaction. Dick says resolution improved along with it.
The same standard applies across the journey. An agent that handles more chats, identifies more accounts, or drafts more advice is useful only if the work improves the result HubSpot cares about.
"The power of AI is not just AI doing everything on its own," Dick says. "It's human authenticity plus AI efficiency."
Dick's advice is to set a high bar for quality and keep people part of the work. In the sales and customer success systems he describes, the assistant supports the employee rather than replacing that person's role: The representative still works the deal, and the customer success manager still spends time with customers.
Dick's starting point for leaders is deliberately simple: Draw the customer journey and write down what the team does to move a customer through it.
Then ask which step creates the most pain. Define the bounded job an agent or assistant could take on. Identify the company context and evaluations it would need. Choose the model that fits that job. Decide which customer or business outcome will determine whether the system is working, and which part of the work people should still do.
HubSpot's representatives showed that building a prospecting agent was not the hard part. The more consequential decision was which version deserved to become part of the company's go-to-market model.
Did you enjoy this transformation story? Go deeper on how AI is reshaping work and business with The Artificial Intelligence Show. Each week, we break down what matters in AI and what leaders should do about it.