The first-hundred test for killing a working AI product

Your AI product works, but the next week is empty. That empty week is the first-hundred test: before scaling, can you name a handful of real users outside your team who used it repeatedly, would pay, and would be upset if it disappeared?

If the answer is no, you have an experiment, not a company.

A working demo is not a customer

DraftSpring was made available as an open-source tool for Ghost blogs, supporting AI-backed ideation, outlines, drafts, edits, image creation, human review gates, and scheduled posting. The article workflow moved each piece from concept and sign-off to outline, draft, human touch, critique, media packaging, final sign-off, and Ghost publication. The initial build included a Python/FastAPI backend, React frontend, SQLite, and magic-link authentication. Its configuration examples listed gpt-4.1 as the OpenAI model ID, gemini-2.5-pro as the Gemini model ID, and claude-sonnet-4-6 as the Anthropic model ID.

The hosted SaaS was shut down, and the repository was cleaned up so Ghost users and technical teams could run or fork the publishing workflow. Competitor Tracker became the project receiving most of the team's attention. A clean pipeline, a useful stack, and a public repository do not create a customer. They only make the product easier to inspect. If nobody outside the team keeps using it, the inspection is a post-mortem, not a business.

Retention is the honest signal

For month-to-month products, staying power shows when customers remain after Month 3, because monthly customers can leave easily. Month 3 is when the core customer base remains after short-lived signups churn. If retention keeps falling after Month 3 and never flattens, early enthusiasm has not turned into lasting usefulness.

High benchmark scores alone do not guarantee dependable performance in real use or sound decisions. A model can produce a clean demo while the user still has to fix the output manually. The signal shows up in the follow-up week, the next week, and the moment the user has to do something without you.

Do the unscalable work to see that signal. Call the users. Sit with them. Watch them struggle. Ask what they did recently. Ask what they would pay. If they do not use it again, stop pretending the pipeline is the product. The unscalable work is the only way to separate a demo from a habit.

Scale numbers do not replace repeat users

The median enterprise AI startup reached $2.1 million in annual recurring revenue (ARR) by month 12, over twice the earlier best-in-class benchmark of $1 million. Bottom-quartile VC-backed AI business-to-business (B2B) companies reached $1.2 million in ARR by year-end, while top-quartile companies reached $5.3 million, a 4.4x spread. Jasper AI reported more than 100,000 active users and more than 850 enterprise clients by 2024. Jasper's scale shows what a repeat base can look like, but it does not tell a founder with no repeat users how to get there.

Those numbers describe companies that already have a repeat base. Growth tactics that need scale are useless before you have a handful of repeat users. You cannot buy a customer base with a launch, a thread, or a paid funnel if the following week is empty. The test forces you to look at that week before the scale numbers make the mistake bigger.

Kill or park before scaling

Killing a working product feels wasteful, but a working product with no repeat users is a more expensive waste. A good product can still fail when the niche is too narrow or founder attention drifts. Write down the names of the people who used it repeatedly on a page. If the list is empty, kill or park the product. If the list is real, protect their time and make the product easier to use.