Review & Confirm

Customers took an average of 14 days to reach first value, and 16% of paying customers left during setup. At roughly $8,000 in lifetime value per activated customer, the onboarding experience was not a cosmetic problem—it was a material growth problem.

Earlier responses—templates, longer wizards, guided tours and documentation—still left the customer doing the configuration. Research made the underlying need clearer: people had bought HubSpot to do their job, not to complete a CRM curriculum first.

As lead designer, I drove the concept and the end-to-end experience. I ran customer interviews, synthesised earlier testing and CSAT patterns, spent three weeks mapping technical feasibility with engineering, and aligned product, engineering and leadership around an MVP.

• End-to-end product design from concept through implementation
• Customer research and synthesis of existing feedback
• AI capability mapping and MVP definition with engineering
• Three major design and testing iterations

The team could use a company website and business context to preconfigure sales pipelines, email templates, contact properties, basic workflows and dashboard reports. AI could not safely connect a website, import customer data, add team members, authenticate email or configure payments.

That boundary shaped the first release. We chose three high-impact, reversible jobs—sales pipeline stages, five tailored sales emails and a reporting dashboard—and made the generated configuration read-only. Editing and regeneration would come after we had proved that customers understood and trusted the model.

The first iteration placed AI inside the existing wizard. Testing showed that automation did not make the wizard feel less like homework. The second iteration displayed every generated configuration at once; people found it interesting, but too dense to evaluate confidently.

The final model separated the work into focused, single-task review views. Each view explained what was created and why, showed a simplified preview, and gave customers a clear choice to accept or skip. The result preserved control without returning the setup work to the customer.

A sales pipeline and a reporting dashboard are both single generated outputs, but they need different kinds of evidence. The pipeline preview kept every proposed stage visible before creation; the dashboard reduced a dense configuration to the reports and sample data a customer actually needed to evaluate.

Both used the same interaction anatomy: explain what AI prepared, show enough of the result to support a decision, and make accept or skip unambiguous. That consistency made the review model easier to learn without flattening the products into the same preview.

Email setup created five separate templates, so a single accept-or-skip decision was not enough. Customers needed to inspect the set, compare useful options and choose only the messages they wanted before anything was saved.

The multi-output version kept selection and preview in one place. It preserved the focused rhythm of Review & Confirm while giving customers control at the individual-item level instead of sending them back through a setup wizard.

Success was not a decorative confirmation. It connected the newly created asset to the next useful action, while also making it clear that the AI setup was complete. Skipping needed equal care: the choice remained reversible and customers could go directly to manual configuration when that was the better route.

I documented review, success and skipped states as a reusable system so teams beyond onboarding could adopt the same trust-and-control model without redesigning it from the beginning.

Time to activation fell from 14 to 9 days. Completion increased by 5%, and 68% of customers accepted the AI-generated configuration.

The next phase focused on editing and regeneration, additional product surfaces and entry points, and exploring non-modal versions of the same review pattern.