Let an AI agent run a SaaS alone, then audit it: what an autonomous build got wrong

Founders tempted to hand a whole product build to an AI agent need to know what goes wrong when nobody sets direction.

The author ran an experiment in which Claude was given a small budget and full autonomy to build a pricing-intelligence SaaS, with no human writing code. The agent produced a large site with a pricing database, calculators, sign-up, Stripe payments and email. The author then audited it and fixed problems. Most individual features worked, but the product had grown into five products, ran several lifetime and monthly deals at once, some purchases did not unlock access, and content presented modelled estimates as real case studies. Fixes: narrow to a one-sentence product identity, cut pricing to Free plus one paid tier, label modelled content with assumptions and dates, standardise navigation, repair a sign-up bug and create a missing Stripe price. The lesson drawn is to add a human decision layer on top of agent output.

Individual features mostly worked, but the whole was incoherent until a human narrowed the scope, simplified pricing and fixed bugs. The post gives no traffic, revenue or user numbers.

One agent in one experiment, judged by the author, who also runs the experiment. The audit and fixes were partly human judgment, so the post cannot say what an agent alone achieves. No usage or revenue data. The specific Claude model is not named. Year of posting is not shown on the page.

Access to an AI coding agent, a Stripe account and an auth provider if you want to repeat the setup.

• DEV Community (Joske Vermeulen) ↗ Firsthand write-up, publication date unknown

Source published: unknown. Last reviewed here: October 18, 2026. Spot a mistake? Tell us.