Home / Blog
Blog
Practical notes for AI teams on getting real-world training data: how licensing works, how de-identification works, and where models break on messy real data.
- How to license real-world training data for AI
A step-by-step walkthrough for AI teams licensing private, real-world data: writing the spec, running a pilot, reading the dataset card, and negotiating license terms and pricing. - De-identifying clinical notes for AI training: what HIPAA Safe Harbor requires
HIPAA's Safe Harbor method lists 18 identifiers that must be removed before health data counts as de-identified. Here is what that means for free-text clinical notes used in AI training, and when Expert Determination fits better. - Code-mixed speech data: why ASR breaks when speakers switch languages
Many speakers switch languages mid-sentence. Here is why speech recognition models struggle with code-mixed audio, what good code-mixed training data looks like, and how to collect it with consent. - Handwritten invoices and bills: the training data gap in document AI
Document AI models handle clean printed invoices well and struggle with handwritten bills, phone photos and crossed-out lines. Here is why, what labels matter, and how to build a dataset of real ones. - Synthetic or real-world training data? Where each one works
Synthetic data is fast and cheap; real-world data is slow and messy. A practical look at where each works for AI training and evaluation, and why the two work best together. - A dataset card template for licensed AI training data
A free, copyable dataset card template for licensed AI training data, covering source, consent, collection, labelling, de-identification, license terms and known gaps. - Where to get AI training data that isn't on the public web
The main routes to non-public AI training data compared: data marketplaces, annotation vendors, custom collection, direct partnerships and your own customers' data. What each is good for, and the questions to ask. - Data provenance for AI training: what it is and how to document it
Data provenance is the documented origin and history of every record in a training set. What it covers, why buyers and regulators ask for it, and a practical way to record it per record. - Retail POS and inventory data for demand forecasting models
What makes point-of-sale and inventory data useful for training demand forecasting and retail AI models: granularity, stock-outs, promotions, store diversity and how to license it from independent retailers. - How AI training data is priced
The common pricing models for licensed AI training data (per record, per audio hour, per project, subscriptions) and the factors that move the price: exclusivity, labelling depth, source access and de-identification. - Selling a shut-down startup's code and data to AI labs
What founders of shut-down startups can license to AI labs (codebases, Slack, tickets, CRM, docs, financials), who controls the IP after a shutdown, what to strip out, and how the process works. - How to license your company's data to AI companies, step by step
A step-by-step guide for companies licensing internal data to AI labs: what to inventory, how to confirm you have the rights, what to leave out, what goes in the licence, and how data is cleaned and delivered. - What is your startup's codebase worth to AI labs?
What makes a private codebase valuable to AI labs for training and evaluating coding agents: never-public code, real git history, reviews, tests, context around the code, clean ownership and the licence terms. - Selling data and IP when a startup shuts down: what founders should know
Who can sell or license a startup's code and data during a shutdown, and which approvals apply: board and investor consent, ABCs, bankruptcy, administration and liquidation in the US, UK and EU. - Is it legal to sell employee Slack and email data for AI training?
What the GDPR, UK GDPR, CCPA and FTC guidance mean for companies licensing workplace Slack, Teams and email to AI labs: lawful basis, employee consent, transparency, anonymisation versus pseudonymisation, and a careful process.
New posts are also available by RSS.
Tell us what your models are missing
A few lines is enough: the domain, the format, and roughly how much.
Request data →