Building and selling a 30-hour niche dataset of voice agent workflows (ElevenLabs)
Hey,
I’ve got an idea that on paper seems like an easy way to make a decent amount of money, but I’m not sure how realistic it is.
My plan is to sell datasets of myself building voice agents on ElevenLabs. The dataset would include screen recordings, mouse movements, tab switches, and full annotations.
The technical barrier to capture the screen and interaction data is quite low, and I plan to outsource the video annotation process to a team in Madagascar.
Having already built live voice agents, I know that building them involves a lot of live iteration and quick fixes, which makes this specific sequential data highly valuable for "computer use" AI models.
My strategy: I want to create a 2-hour high-quality sample to showcase on X and Reddit, and use it to cold-pitch potential buyers via email and LinkedIn DMs. If there's proof of interest, I can easily produce a 30-hour dataset within a week or two. Even though it’s not thousands of hours, I believe this specific type of workflow data is rare enough that AI labs might buy it.
My three main questions are:
- Is this kind of niche workflow dataset interesting enough for labs or developers to actually buy?
- What's the best way to get this in front of the people who have the budget for it?
- Is a 30-hour high-quality dataset for a complex task substantial enough to be sold standalone?
I got this idea after seeing Markov in the current YC batch. The data samples on their website look high-quality but totally feasible to replicate for specific niches.
If anyone has thoughts or experience in the AI training data market, I'd love to hear them!