AI-First Onboarding

HubSpot's Pro tier is where most of the company's revenue sits, and getting customers up and running is part of what they pay for. In 2024, fixing Pro onboarding was the company's #1 product priority.

The deeper problem: completing setup wasn't the same as getting value, and for a large segment of customers there was no path to value at all except a paid onboarding specialist.

Finding the real problem

I isolated the cohort that made the problem measurable: customers who had waived human-led onboarding entirely. Because no person was involved, any gap in their activation was down to the product, not a human. That comparison was the proof - purely product-led customers activated meaningfully worse than human-onboarded ones.

It did two things at once. It quantified the gap a product-led experience had to close, and it gave us a clean group to build and test against - with no human onboarding to confound the results.

The before state

Two types of customers landed in onboarding, both underserved.

Customers who upgraded without speaking to sales saw a generic task list - the same one regardless of which hub they'd bought or why. A sales rep who had just upgraded to Sales Pro was told to “import contacts, create properties, invite teammates.” Generic setup advice with no connection to why they upgraded or what they were trying to do.

Customers who bought through a sales rep and waived human-led onboarding had it worse: no guided experience at all. They'd closed a deal and had no one to walk them through the product they'd committed to.

The data matched the experience: 33% of customers accepted an onboarding plan and engaged with it. Two-thirds ignored it.

What good actually looked like

Before writing a spec, I ran customer interviews and sat in on workshops with our human onboarding team - the people who do this every day for enterprise customers.

The pattern was clear: good onboarders don't start with a task list. They start with questions. What are you trying to accomplish? Who on your team will use this? Have you used a CRM before? The task list is a consequence of the answers, not a template applied to everyone.

The hypothesis: if we could capture those goals upfront - role, company size, hub purchased, prior CRM experience - we could generate a plan that felt relevant rather than generic. An AI agent could do the intake that a human onboarder does, at scale.

Building to learn

Before engineering wrote a line of production code, I built a prototype in Lovable covering the full journey: upgrade, conversational intake, plan acceptance, first task.

Customers didn't want to chat their way through onboarding. They wanted to click. What resonated wasn't the conversation - it was the output. When they saw a plan that reflected their specific goals and the hub they'd bought, they engaged. The conversation was a means to an end, not the experience itself.

This shifted the team's direction before any production code existed. We moved from “build a conversational agent” to “build an agentic intake that produces a personalised plan with a familiar UI.” The agent's job was to understand the customer, not to talk to them.

The agent, and how I knew it was any good

The experience opens with a structured intake. The agent reads the customer's website and business context, pre-fills a profile, and asks them to confirm or edit it - a review-and-confirm step rather than a questionnaire. A few targeted questions about goals and team setup follow. The agent uses those answers, plus company size and industry, to generate a personalised plan from a library of onboarding tasks and tours.

The hard part wasn't generating a plan. It was knowing whether the plan was any good. So I built the yardstick. I analysed 100 real calls from HubSpot's human onboarding specialists - the people who do this well every day - and turned them into a golden eval set defining what a good plan looks like for a given customer profile.

I wrote the first-pass system prompt the agent runs on, owned the task and tour library it draws from, and ran the agent's outputs against those evals - iterating the prompt with engineering until the agent's plans matched what a specialist would have recommended for the same customer. The expertise that used to live in a paid human became the bar the agent had to clear.

Results

The onboarding experience I owned - intake, plan generation, and routing, built with partner teams who shipped the rendering surface and task content - drove a +20pt lift in week-1 Sales Workspace visits and +26pts in month-1 usage against a control. $270K of MRR reached value through this product-led path, work that previously required a paid onboarding specialist. Portal Initial Value - whether customers reach meaningful usage in their first 30 days - rose from 56% to 69% across the period. The framework expanded to Marketing, Service, Commerce, and Content Hub.