The $43K SaaS replacement that shifts the hard part to maintenance

The contest. The organizer of the AI Engineer conferences published a 'Kill my SaaS' challenge after SessionBoard quoted $43,000 a year for four events. 72 teams built replacements from a written brief, 42 screenshots, and a walkthrough video. Five apps cleared the bar for real adoption, and the winning entry from Valentin shipped multiple portals and a schedule the events team wanted to use. The estimated build cost: roughly $1,000 in tokens plus about $10,000 in AI engineering time.

The workflow. None of the finalists typed a prompt and walked away. They researched before coding: one crawled SessionBoard's help portal with Firecrawl, one fed the walkthrough video and API docs into Claude, and one mapped every feature flow before launching parallel agents. Planning was most of the work. The prompt came last. That suggests agents are useful for execution, but the quality depends on the human's knowledge of the product and domain.

The evals. Real users disagree with automated scoring. The team built an AI judge that had Claude log in and click through 20 scenarios; it was useful for bootstrapping, but its top 10 included six apps that human reviewers rated weak. The eventual winner ranked 19th on the AI leaderboard. Humans caught what agents missed: multiple portal types, drag-and-drop schedule state, and adding a talk without going through a CFP form.

The pattern. The builder workflows are the most transferable signal. Claude Code and Codex were roughly tied, and most teams used both. Three patterns emerged: a stronger model plans then a cheaper model writes; one model builds and another checks; no model approves its own work. One team ran Claude and Codex in separate workspaces and required one to verify the other before merging. Separate verification is an operational discipline that matters more than any model feature.

The trade-off. The hard part starts after launch. Someone must fix bugs at 2 a.m., handle integrations, and keep the system running, a real shift from vendor support. The transcript includes a sponsor with a ticket bug and a support chain ending in 'refresh your cache'; the company decided it wanted software it could actually own. But owning means staffing an internal software team, even if some of it is designers and operations people directing agents.

The verdict. This is not proof that AI killed SaaS. It is a credible, transparent experiment showing that a mid-tier tool can be replaced when a team has real user requirements, agent orchestration skills, and an appetite for maintenance. The $43K savings is real, but the long-term cost of ownership is easy to underestimate. Engineering leaders should treat this as a bounded pilot option, not a mandate to stop buying software.