AI labs are organisations or teams that run experiments to develop and assess artificial intelligence systems. They might build a model, test an existing one or investigate why it fails on a particular task. The phrase AI data labs is less precise: it could describe researchers, a data preparation service or a marketing label. Ask what work is actually done.
For a UK business considering a licence, the distinction isn't academic. Many teams use machine learning, but the organisation requesting records might be the developer, a supplier or someone arranging an introduction. ADP (Accord Data Partners) is independent of buyers and makes introductions, rather than building models or taking company datasets. Any commercial agreement is for the business and buyer to make.
What makes an organisation an AI lab?
AI labs are defined more by their work than their name. A university group, a company unit or a dedicated research organisation can all qualify if they frame questions, run controlled experiments and assess results. An independent introducer, by contrast, can connect parties without designing tests or training systems. Job titles alone won't tell you which role an organisation plays.
A model developer decides what the system should predict or generate, which inputs it can use and what counts as an acceptable output. That may involve writing software, choosing an existing model to adapt and checking mistakes. A firm that only packages files for someone else isn't necessarily doing that development work, even if its name includes ‘lab’.
Look for the decisions a research team controls. Can it change a model, choose how results are measured and investigate failures? If it merely cleans records to a specification supplied by another party, it's doing useful preparation but not necessarily the experiment. Some organisations do both, so ask which entity sets the purpose for the proposed use.
What do researchers work on?
AI labs don't spend every day creating a brand new chatbot. Researchers may try to make a system read a particular document, spot mismatched fields or give more reliable answers. A project starts with a narrow question and a way to check progress. Without that, a folder of interesting records may have no clear use in an experiment.
Researchers then decide what counts as a useful input and a trustworthy answer. Training examples for document classification might pair a document with the category a reviewer has checked. Google's guide to dataset labels explains the difference between a direct label and a proxy. If the category is inferred from a loose shortcut, the model may learn that shortcut instead.
AI labs also spend time finding where a system breaks. They might compare two prompts, review difficult outputs or check whether performance changes when document layouts vary. That's experimental work, not evidence that every record has commercial value. The right description of a project is its intended task, inputs and checks, not simply ‘we use AI’.
How do models move from training to testing?
A typical workflow separates material used to adjust a model from material reserved for checking it. An evaluation set helps show how the system handles examples it wasn't allowed to learn from during that round of training. If the same document or near duplicate appears on both sides, a convincing score may reflect recognition rather than a useful skill.
A benchmark is a repeatable test, not a promise that a system will work in every setting. Teams choose measures that reflect the task: overlooking a costly error may matter more than occasionally flagging a harmless one. Google's machine learning glossary describes why accuracy can mislead when categories are uneven, and explains measures such as precision and recall.
AI labs may repeat this cycle: review mistakes, change the approach and test again against suitable untouched material. A better score on one test isn't the whole decision. Teams still need to check how the system responds to unfamiliar formats and whether its outputs can be reviewed by the people who would use it. Testing narrows a claim; it doesn't remove uncertainty.
Why do specialised teams need different examples?
Hypothetical worked example: A repair firm has service notes marked ‘warranty’, ‘chargeable’ or ‘unclear’. A team wants to suggest a category, not decide what a customer owes. It first defines what counts as a warranty visit, including exceptions. That definition shapes the model behaviour being tested; collecting a large pile of unmarked notes wouldn't answer the same question.
AI labs tackling that example would want variation that matches the decision: different wording for return visits, notes that mix faults and invoices, and genuinely unclear cases. They would also keep similar notes from the same job together when dividing material for testing. Otherwise, a system might seem capable because it has already seen nearly identical wording.
A specialist task needs examples with relevant distinctions, not just more files. A technician might need to resolve borderline labels that an outside annotator can't judge reliably. If the source notes contain names or free text, the preparation question also changes. The owner would first have to decide what, if anything, can be shared for the stated purpose.
How does a lab differ from a data supplier?
A data supplier gathers, prepares or licences material for another organisation to use. Its work may include checking formats and documenting where records came from. That doesn't mean it owns every right in what it handles or sets the experiment's objective. If someone calls itself a supplier, ask whose records are involved and what authority it has to provide them.
AI labs may buy access to information, assemble material internally or work with someone who arranges introductions. Those are different roles in a transaction. ADP is not a lab or a destination for company datasets: it is independent of buyers and helps parties make contact. The data owner and buyer must agree their own licence, if they decide to proceed.
An intermediary might pass on a description of an opportunity without receiving the records. Ask whether it introduces parties, negotiates terms, hosts files or processes information; these are different activities. An introduction isn't approval to use the material. Before agreeing terms, see how AI training data licensing works, then check who would have access under the proposed arrangement.
Is an AI data lab a buyer?
AI labs might be called AI data labs, but the second phrase doesn't reliably identify a buyer. It could mean a team that labels information, a research group or a service selling software to other businesses. A name on a website can't tell you whether that organisation wants a licence, acts for someone else or simply advertises technical services.
Instead, ask who sets the research question and who would actually use the records. If the contact can't explain the intended task or the entity that would receive permission, don't assume the lab named in an email is the counterparty. A sensible first reply seeks roles and proposed uses, not a sample export or an assurance that the records are valuable.
The requested use also matters. Held-out examples for testing serve a different purpose from records used to change a model, even when the file format looks identical. Ask whether material would be used for training, evaluation or both, who could inspect it and whether onward use is proposed. Those answers are more informative than the label attached to the organisation.
What does an owner need to ask?
If your company is approached by AI labs, start with a description, not an export: types of records, the period covered and why your business holds them. Then ask who would receive them, which task is proposed and what the requested access includes. A useful discussion can begin with categories and boundaries without exposing real customer entries or confidential details.
Next, ask what happens between testing and model deployment. Would the buyer keep copies, retain derived outputs or use the material to improve later systems? Get the permitted purposes, recipients, security arrangements, retention and deletion process in writing before deciding whether any licence is suitable. A permission for one experiment shouldn't silently become permission for every later use.
AI labs can't cure an ownership or privacy problem by offering to sign a licence. Business consent to an introduction is separate from a lawful basis for processing personal data and from confidentiality, copyright and contract authority. Pseudonymised records may still be personal data. Use the records licensing safeguards guide to organise those separate checks, and seek appropriate professional advice if needed.
How ADP fits into the chain
ADP's job sits before data sourcing or licensing decisions between the parties. It offers a paid introduction course, a portal and resources for making business introductions; it doesn't train models or receive company datasets. That distinction lets you ask one practical question first: do you want to learn how introductions work, or are you trying to assess records you already control?
For an owner, licensing scope belongs in direct discussions with a potential buyer, not in a course enrolment. Define the specific records, purpose, permitted users and limits you would consider, then have the buyer's terms reviewed. ADP can't approve eligibility, grant a licence or promise earnings. Even an agreed licence doesn't by itself settle privacy obligations.
If you're a business owner seeking clarity on the introduction role, read ADP's guide to AI data brokering. It explains the route without asking you to send records to ADP. You can then decide whether an introduction is relevant, before spending time on a possible licence or assuming that a research organisation will want your material.
Frequently asked questions
Do all AI labs train their own models?
No. Some adapt existing models, run evaluations or investigate failures. Ask which parts of development the organisation actually controls.
Does a lab name mean an organisation wants to buy my records?
No. It may do research, prepare data or sell technical services. Ask who would use the records, for what purpose and under whose licence.
Can pseudonymised records be licensed without a privacy review?
Not automatically. Pseudonymised information can still be personal data, and a licence doesn't replace privacy, confidentiality or rights checks.
