Four years is long enough to separate a good AI bet from a lucky one.
In late 2022, two events compressed a decade of AI adoption into about six weeks. Stable Diffusion went open-source in August, putting image generation in the hands of anyone with a GPU. Then ChatGPT launched in November and added 1 million users in five days. The window for staking a position in AI was open, and founders and investors rushed through it.
Now that enough time has passed, patterns are emerging around what actually held up.
The Wrapper Problem Claimed Most of the Early Winners
The first wave of 2022 AI startups was largely built on a simple premise: take GPT-3 output, add a clean interface and a specific use case (copywriting, customer support, email drafting), and charge a subscription. This worked well for about 18 months.
Jasper AI is the cautionary example most people cite. The company raised $125 million at a $1.5 billion valuation in October 2022, right before ChatGPT launched. Within a year, GPT-4 had made the underlying capability so cheap and accessible that Jasper's value proposition collapsed. The company laid off 30% of its staff in 2023. By 2024, it had pivoted toward enterprise marketing platforms to find a more defensible position.
Jasper was not alone. Dozens of "AI-powered writing tools" from that era quietly shut down or folded into competitors after OpenAI's improvements erased the gap between their product and a free ChatGPT subscription.
What Survived Had More Than a Model Behind It
The 2022 bets that are still standing tend to share a few traits. They built proprietary data pipelines their competitors couldn't replicate. They embedded deeply into existing workflows rather than asking users to adopt a new one. Or they picked a vertical narrow enough that a general-purpose model couldn't replace their domain-specific tuning.
Legal AI, medical AI, and code-specific tools held up better than generic content tools. That's partly because regulated industries move slowly (giving companies time to build moats) and partly because the cost of errors in those domains keeps users from trusting a one-size-fits-all model.
Companies that bet on infrastructure rather than applications also came out ahead. Vector database companies, fine-tuning (the process of training a model on your specific data to improve it for your use case) platforms, and AI observability tools all grew steadily as more businesses tried to build their own AI products and needed the plumbing to do it.
The Honest Lesson
The founders who are most candid about this era tend to say the same thing: they underestimated how fast the foundation models themselves would improve. A startup that took six months to build a feature in 2022 often found that OpenAI or Anthropic had shipped the same capability natively by 2023.
The bets that worked were the ones that built on top of AI capabilities rather than around them - solving a real workflow problem for a specific customer type, rather than just making a model's output prettier. That lesson applies just as much to anyone building AI products today as it did in 2022.