Testing a Checkout With Synthetic Customers
Before Wellhound's paid beta, I had a problem every solo founder has: not enough real users to test with, and not enough time to wait for them.
So I built AI customer-testing agents.
Grounded in real people
The agents were not generic "act like a user" prompts. Each one was grounded in real dog-owner reviews and interview patterns: the anxious senior-dog owner, the price-sensitive new owner, the skeptic who has never trusted an at-home test. Each had a goal (sign up, book a draw, understand a result) and the patience level and objections of the person it was based on.
What they did
The agents walked the real product end to end: landing page, signup, dog profile, plan selection, checkout, booking, and results. They logged every point of confusion, every dead end, and every place they would have quit.
Results
• 96 pre-launch defects found, from broken states to unclear copy to missing edge cases
• Simulated checkout completion improved by 20 percentage points after fixes
Where it fits
Synthetic testing does not replace talking to customers. It replaces the part of QA where you, the founder, click through your own product and cannot see it anymore. It finds friction. It cannot tell you whether someone will pay. That is what the real beta is for.
What I took from it
• Ground test agents in real customer language, or they will just confirm your assumptions
• Use AI to find friction cheaply, so real users spend their attention on whether the product is worth it