Build and sell licensed datasets to AI companies
AI teams need niche data the open web doesn't have, with rights they can defend to a lawyer. Pick a niche nobody has packaged, collect it with written rights, clean it to a buyer's standard and sell it as a licence you can grant again and again. Budget for contributor pay, a solicitor's review and insurance before your first sale.
- A shop window: a gated sample on Hugging Face plus the data-partner routes into marketplaces
- A kit: contributor agreement and licence outlines, a pricing calculator, outreach messages and a 30-day plan
- Python and uv, from Put AI in your terminal. The pipeline installs its one dependency itself
- Comfort with the terminal: Terminal survival kit
- Access to a niche: a trade, a profession, a language, a place or a kind of expertise you can recruit contributors from
- Working capital for contributor pay before any sale. The kit's example uses about £13,000 for version 1
- A business set up properly: Make it legal
Know what AI buyers actually pay for
Buyers don't pay for rows. They pay for data they can't get anywhere else and can prove they're allowed to use.
The web is not a dataset you can sellEvery serious lab already has the public web. What it lacks is data that is niche (a domain the web covers badly), clean (deduplicated, consistent, no junk), documented (a card that says exactly how it was made) and licensed (a signed chain of rights from the person who created each row to the buyer). Miss any one of the four and a data team can't use it, however big it is.
Rights are the productModel makers now have to account for where their training data came from, so a data team's lawyer asks one question first: where did each row come from, and who signed what? A clean answer to that is worth more than the rows, and it's exactly what this mission builds.
What sellsLook for gaps where the web is thin, private or messy:
- Expert Q&A and reasoning in a trade or profession: UK plumbing, veterinary nursing, conveyancing steps, agronomy, aircraft maintenance logs (with permission). Written by people who do the job.
- Speech in under-served accents and languages: regional UK English, Scots, Welsh, British Sign Language video, code-switching between English and a community language.
- Images of things the web doesn't photograph well: roof defects, damp, wiring faults, crop disease in UK fields, worn machine parts, handwritten forms.
- Evaluation sets: a few hundred hard, expertly checked questions with answers that nobody has published. Teams pay well for a test the model has never seen.
- Preference and correction data: experts rating or fixing model answers in their field.
Who buysNot only the big labs. Startups fine-tuning a domain assistant, enterprise AI teams building internal tools, speech and vision companies, evaluation firms, and marketplaces that resell licences. Smaller teams decide faster and are easier to reach.
Careful- What it takes. Capital first: you pay contributors before you sell anything. The kit's example version 1 costs about £13,000 including a solicitor's review and insurance.
- Time: 10 to 15 hours a week for the first three months: recruiting, reviewing, building, selling.
- Skills: basic Python (the kit does the heavy lifting), clear writing, and enough knowledge of your niche to judge quality.
- Paperwork: a written contributor agreement reviewed by a solicitor, a licence template, and professional indemnity insurance that covers IP claims.
- Time to first customer: usually two to four months from first contributor, because data teams test a sample before they buy.
- Who this suits: people with a foot in a trade, profession or community the AI world can't reach, who like building systems and can sell to technical buyers. If you already manage freelancers or run a small agency, you have most of the muscles.
Everything you'll use is in one kitDownload the kit (16KB): the sample generator, the pipeline, the agreement and licence outlines, the pricing calculator, outreach messages and the scoreboard.