← Back to blog

AI Training Data: What It Is, Where It Comes From, and How to License Yours

Company records sorted for AI data licensing

AI training data is the collection of examples a model learns from: text, images, audio, code or records. Stanford HAI describes it as the examples used to teach machine learning models to perform specific tasks. Developers get it by collecting it, scraping the web, hiring contractors, buying from data brokers or licensing it from the people who hold it. If you run a company, you may already hold the kind of data they want: emails, support tickets, CRM history and documents built up over years of real work.

This guide covers what training data consists of, how developers find and procure it, what separates good data from bad, and what to check before you license your own.

Nordicmartech
Put Your Operational Data to Work
Mode connects companies holding emails, support tickets, and CRM history with AI developers seeking real-world training data, while you retain ownership.
Explore data licensing

TL;DR

Table of Contents

What AI training data actually is

A model has no knowledge until it sees examples. Training data is those examples. For a language model they are passages of text. For an image classifier they are pictures. For a customer-service assistant they might be thousands of real conversations between customers and agents.

According to IBM, training data consists of features or attributes. In supervised learning the data is usually labeled, so the model knows the correct output for each example. In unsupervised learning the data is unlabeled and the model looks for patterns on its own.

In practice that gives you two broad categories:

Examples of training data

The format varies by task. IBM lists spreadsheets, video footage, web pages and PDFs among the forms training data can take. For business use, the common examples look like this:

Data type What a model can learn from it
Support tickets How customers describe problems and how agents resolve them
Email threads Professional tone, negotiation, scheduling, follow-up patterns
Chat transcripts Short-turn dialogue, intent, escalation
CRM histories Sales cycles, account progression, customer signals
Documents and contracts Structure, terminology, domain-specific language
Code repositories Programming patterns, review comments, bug fixes

Public web pages cannot supply most of these. They sit inside companies, which is why developers look for licensing arrangements.

How training data is split: training, evaluation and testing

Not every record is used to teach the model. Microsoft’s guidance explains that data is divided into training, evaluation and testing sets. Evaluation data helps monitor performance during training and tune hyperparameters. Testing data checks predictive power on records the model has never seen.

You will see a ‘30% rule’ mentioned in search results. There is no single standard behind that phrase. The split we can point to is a common rule of thumb from the University of Cambridge’s AI core concepts: 80% of the data for training, 10% for development and 10% for testing. Treat it as a starting point, not a law. The right split depends on how much data you have and how the model will be used.

Proportional split of training, development, and testing data

Where AI developers get training data

The U.S. Copyright Office’s report on generative AI training, in its pre-publication version from May 2025, describes the main routes: collecting data from user interactions, scraping the web, hiring contractors, purchasing from data brokers, and licensing directly from rightsholders. The FTC’s staff report on cloud and AI partnerships lists the same set for the U.S. market and adds that strategic partnerships with cloud providers can allow training data to be shared.

Each route has a different risk profile:

  1. User interactions. Convenient for companies that run a product, but the data is limited to what their own users generate.
  2. Web scraping. Cheap and broad. Rights, consent and quality are hard to document.
  3. Contractors. Good for creating custom, labeled examples. Slow and expensive to scale.
  4. Data brokers. Fast to buy, but the chain back to the original source can be thin.
  5. Direct licensing. The holder grants defined rights under a signed agreement. Provenance is clear, and that clarity is what buyers increasingly ask for.

If you are the buyer: finding and procuring data

Developers searching for training data, whether for acquisition or procurement, should ask the same questions of any source. Who created the data? What rights does the seller hold? Was personal information removed, and how? Is the data representative of the task? A licensed archive of real operational records answers those questions better than most scraped collections, because there is a named company behind it and a contract that defines what you may do.

What makes training data high quality

Volume gets the attention. Quality does the work. The Copyright Office report describes high-quality training data as the product of rigorous, iterative curation: cleaning, filtering out low-quality content and deduplication, together with documentation of provenance and metadata.

A U.S. Department of Health and Human Services study on approaches to developing AI training datasets points to predefined criteria such as completeness, consistency and representativeness. It also notes that proprietary datasets currently dominate AI development, which limits transparency and reproducibility, and it argues for more transparent development practices.

A NIST presentation on AI in software supply chains lists what dataset quality touches: bias, hallucinations, regulatory compliance, license and IP compliance, rights management and data sovereignty. Data quality is not only a modeling concern. It is a legal and operational one.

A practical quality checklist

Operational records often score well on the third item. Real customers write messy, varied, specific messages, which is what a model needs to handle real use.

Why provenance and rights decide the deal

A large-scale audit published in Nature Machine Intelligence traced more than 1,800 text datasets and found license omission rates above 70% and license error rates above 50%. The same audit points to tools such as the Data Provenance Explorer for tracing data and filtering datasets by license conditions. If public datasets are that hard to trace, a clean, documented, properly licensed archive stands out.

The FTC has also shown that how data was gathered matters. In a staff report on generative AI and the creative economy, the agency recounts its allegation that Amazon and Ring used Alexa voice recordings and Ring camera videos to train algorithms while violating customers’ privacy. The lesson for any company holding customer records: the permission to use data for one purpose does not automatically cover AI training.

Document before you license

A data governance plan written early makes every later step easier. A comment submitted to NIST by data governance specialist Kathy Rondon recommends creating one during the design phase. It should record the training data, how it was collected, its provenance and lineage, PII protections, sharing and retention policies, handling procedures and who is accountable. The same comment argues that trust in AI rests on rigorous governance of the underlying data, not on the model code alone.

For a company preparing records for licensing, a useful version of that plan answers six questions:

  1. Which systems hold the records, and what date range do they cover?
  2. Whose information is in them: customers, employees, partners?
  3. What do your customer terms, privacy notices and contracts say about the use of that information?
  4. Which third parties have rights or confidentiality claims over parts of the content?
  5. How will personal and sensitive details be removed or masked before anything leaves your control?
  6. Who inside the company signs off, and who is the contact for the buyer?

General counsel will recognize this as a standard diligence list. Doing it first shortens the path to a signed agreement.

Licensing your own operational data

Most companies think of data as a cost: storage, security, compliance. For AI developers, years of real support tickets, emails, chats, documents, CRM histories and code are scarce material. The data reflects how businesses actually talk and work, in finance, software, consulting and other sectors, and it is hard to reproduce synthetically.

The model we built for this is simple to describe:

We state the potential range for the program as $100K to $5M, depending on the quality and volume of data shared. That is a range for the opportunity, not a promise for any single company. What you could earn depends on what you hold and how clean it is, which is why we start with a readiness review rather than a quote.

What tends to make records more valuable

Using the quality criteria above, these are the traits we look at first:

Perspective

We hear the same assumption from COOs and general counsel: the data is a liability, so the safest move is to leave it alone. We see it differently. The records your company produced while doing its job are an asset that is currently sitting idle, and the AI market is short of exactly this kind of material.

What holds companies back is rarely the idea. It is the fear of losing control. That is why we built the process around ownership. You keep your records. Buyers receive de-identified copies. Nothing is exposed until there is a signed agreement, and a readiness check comes before any commitment.

The public record backs the focus on provenance. Audits of public datasets keep finding missing or wrong license information, and regulators have shown they care how training data was gathered. A clean, documented archive from a company that can show where it came from is the opposite of that. We think buyers will pay for that clarity, and that established companies in finance, software and consulting are well placed to supply it.

If you are weighing this, start by finding out what you hold. It costs little and tells you a lot.

Fredrik

How to find out whether your data fits

The first step is not a legal review or a technical project. It is a short fit check. It looks at what systems you run, how long they have been in use, roughly how much data they contain and what rights questions are likely to come up.

A typical path looks like this:

  1. Inventory. List your record systems: help desk, email, chat, CRM, document stores, code repositories.
  2. Rights review. Have counsel read customer terms, privacy notices and key contracts against the intended use.
  3. Readiness check. Assess completeness, consistency and structure.
  4. De-identification plan. Decide how personal and sensitive details will be removed from copies.
  5. Agreement. Only at this stage do records move toward a buyer, under a signed license.

You can see how we handle the common questions on our Mode questions and answers page, and the Mode terms guide explains the conditions in more detail.

Start with a fit check

If you own or run an established company and want to know whether your operational records could be licensed for AI training, our fit check is the place to start. It takes you through the questions above and tells you where you stand, without any of your data leaving your systems. Teams with several stakeholders can use the fit check for teams, or you can book a 20-minute call to talk it through.

Nordicmartech

If you prefer to begin on your own, take the company data licensing fit check.

Common mistakes when preparing data for AI training

Treating volume as the only measure. A million duplicated, inconsistent records are worth less than a smaller, clean set. Deduplicate and standardize before you assess value.

Skipping the rights check. Customer terms written before AI training existed may not address it. Review them with counsel first, and do not assume that internal use permission extends to licensing.

Confusing de-identification with deletion of a few fields. Names and email addresses are the obvious part. Free-text fields such as ticket bodies and email threads can carry identifying details in the content itself. A real de-identification plan covers unstructured text.

Ignoring third-party content. Emails and documents often include material from partners, vendors or clients. Know whose content is in the archive before you license it. Rightsholder disputes are a live issue in the sector: music publishers, for example, told the Copyright Office that sublicensing copyrighted content to AI developers without the owners’ express permission is direct infringement.

Documenting nothing. If you cannot show where data came from and what was done to it, a careful buyer will not proceed. The governance plan described earlier is the fix.

Common mistakes when preparing data for AI training

FAQ

Where can I find AI training data?

Developers use public and academic datasets, web collections, contractors who create custom examples, data brokers, and direct licenses from companies that hold the records. For real-world business data such as support tickets, emails and CRM history, direct licensing from the company that generated it is the route with the clearest provenance. Whatever the source, ask for documentation of origin and rights.

What does AI training data consist of?

It consists of examples such as text, images, audio, video, spreadsheets and PDFs, according to Stanford HAI and IBM. In supervised learning the examples are labeled with the correct answer. In unsupervised learning they are not.

What is the 30% rule in AI?

There is no single standard rule by that name. The split most often cited is 80% of data for training, 10% for development and 10% for testing, per the University of Cambridge. Teams adjust the proportions to the size of their dataset and the task.

How do I collect data for AI training?

Start by defining the task, then choose sources that match it: your own interaction data, licensed archives, contractors for custom labeling, or public datasets with clear licenses. Clean and deduplicate, keep records of provenance, and split the data into training, evaluation and testing sets. Document everything in a governance plan.

Can my company license its data without giving up ownership?

Yes, under our program you keep ownership of your records and only de-identified copies are shared, and nothing is shared until a formal agreement is reached. Whether your own records qualify depends on their content and the rights attached to them, so the fit check is the right first step.

Sources

Made with BabyLoveGrowth to earn links