Glossary: the terms in a company data deal
Plain definitions of the words you meet when a company licenses its data to an AI lab, and when that data has to be cleaned to a privacy law. Each one links to the page that goes deeper. General information, not legal advice.
A
- AI lab
- A company or research group that trains large AI models. Labs license data from companies because real work records show how tasks are done, which public web text does not. Selling company data to AI labs.
- Anonymization
- Changing data so that people can no longer be identified from it by any means reasonably likely to be used. Most privacy laws treat truly anonymous data as outside their rules, and each sets its own test for when that point is reached. Privacy rules by region.
B
- Broker fee
- What we are paid for arranging a deal. The AI lab pays it, and only after the deal closes. About and how we earn.
C
- Company data
- The records a company creates by doing its work: support tickets, emails, quotes, work orders, project histories, procedures. Owners also call it business records or operating records. What kinds of records can be licensed.
- Consent
- A person agreeing to a use of their information. Whether consent is needed, and what it must cover, depends on the privacy law that applies to the people in the data. Privacy rules by region.
- Covered entity
- Under the US health privacy rule (HIPAA), a health plan, a clearinghouse or a health care provider that bills electronically. Health information held by a covered entity can leave only through a permitted route, such as de-identification to the standard in the rule. United States rules.
D
- Data card
- A one-page description of a dataset: what it contains, how big it is, what period it covers, where it came from, how it was cleaned and what rights go with it. A lab reads it before asking for a sample. How a deal runs.
- Data licensing
- Letting an AI lab use a copy of your data under written terms. You keep ownership and keep using the data yourself. You license data; you do not sell it. Sell or license your data.
- Data room
- A controlled place where an approved sample is reviewed. Access is limited to named people, and the company decides what goes in. How a deal runs.
- Data valuation calculator
- Our free tool that asks a few questions about your records and returns a rough range. It is an estimate, not a final offer: a real offer needs a sample and comes in writing. What is my data worth?.
- Date shifting
- Moving every date in one person's records by the same random number of days, so the order and spacing of events stay but the real dates are hidden. Under the US health rule, dates other than the year are among the identifiers to remove unless an expert approves another method. De-identification.
- De-identification
- Removing or replacing the personal information in data, then testing whether anyone could still be identified. Each privacy law sets its own test, so the work is done to the strictest one that applies. De-identification service.
E
- Egocentric data
- Video recorded from the point of view of the person doing a task, for example from a head or chest camera. Labs look at it to learn how hands-on work is done. It often shows other people, so consent, blurring and audio removal matter. First-person video.
- Eval
- A test an AI system can sit. It has a task, a way to check the result and a score. Good evals are built from real work, which is one reason labs look for company records. What is recursive self-improvement?.
- Exclusive broker agreement
- A written agreement for a fixed period under which one broker presents your data to AI labs. It exists so the broker is not duplicating effort and can keep good relationships with the labs. It is different from an exclusive license to a lab, which is a separate term worked out in the deal. Check an offer.
- Expert Determination
- One of the two ways to de-identify health data under the US health privacy rule. A qualified expert in statistics finds that the risk of identifying anyone is very small and documents the methods and results. We do not make this determination; the expert does. United States rules.
F
- Fine-tuning data
- Examples used to adapt an existing model to a task or an industry. Company records, once cleaned, can serve as fine-tuning examples.
K
- k-anonymity
- A check that every combination of identifying details in a dataset, such as age band, town and job title, is shared by at least k people. If one combination belongs to one person, that person can be picked out. De-identification.
M
- Motivated intruder test
- A re-identification check used by the UK regulator. You ask whether a reasonably competent person with ordinary resources, such as the internet and public records, who is determined to identify people, could do so from the data. United Kingdom rules.
P
- PHI
- Protected health information: health data that identifies a person, held by an organization covered by the US health privacy rule. It leaves a covered organization only by a permitted route. United States rules.
- PII
- Personally identifiable information: anything that points to a person, such as a name, email address, phone number, account number, or a detail that identifies them in combination with others. What counts as PII.
- Perpetual use rights
- A license with no end date for the data the lab has already trained on. A trained model cannot give the data back, so use rights are usually perpetual. Exclusivity is different and is never perpetual. Check an offer.
- Provenance memo
- A short record of where a dataset came from, who owns it, how it was collected and cleaned, and which agreements cover it. It lets a lab and its counsel check the data without asking you to repeat the story. What we will not broker.
- Pseudonymization
- Replacing names and identifiers with codes while a key still links them back. The result is still personal data for whoever can use the key, so codes for names are not enough on their own. Privacy rules by region.
Q
- Quasi-identifier
- A detail that does not name a person but can identify them when combined with others, such as age, postal code and a rare diagnosis, or a job title and a start date. De-identification.
R
- Re-identification
- Working out who a person is from data that was cleaned. A de-identification record shows how the risk was tested and who could attempt it. De-identification.
- Reinforcement learning (RL) environment
- A practice setting where an AI attempts a task, gets a score and tries again. Labs build them from real business workflows: the task, the grader and the recorded attempts. What labs build from workflow data.
- Restricted transfer
- Sending personal data to an organization in another country in a way a privacy law limits. The UK and the EU require a transfer tool such as an adequacy decision or approved contract terms; other regions have their own rules. Privacy rules by region.
- Rights-clean data
- Data the company owns or may license, with no third-party claim and no contract that bars the use. It is checked before a sample is shown. What we will not broker.
S
- Safe Harbor
- One of the two ways to de-identify health data under the US health privacy rule. It removes a list of 18 kinds of identifiers, such as names, small geographic areas, most dates, phone numbers and record numbers, and requires that there is no actual knowledge the rest could identify someone. United States rules.
- Sample
- A small slice of your data, de-identified and approved by you before it leaves, that a lab reviews to judge the whole set. No one can price company data without seeing a sample. How a deal runs.
- Special category data
- Personal data that privacy laws such as the UK and EU GDPR treat as more sensitive: health, racial or ethnic origin, political opinions, religion, union membership, genetic and biometric data used to identify someone, and sex life or orientation. Free text in tickets and emails can contain it by accident. European Union rules.
T
- Trajectory
- The recorded sequence of steps an agent or a person took on a task, with the result. It shows how the work was done, not only the answer. Sell workflows.
- Training data
- The examples a model learns from. For company data, the useful part is usually the chain from a request to an outcome, not an isolated file. What kinds of records can be licensed.
W
- Workflow data
- How a team gets work done, captured as documented processes, recorded task runs or AI-assisted tasks. It is priced per workflow, and what is captured is agreed with you first. Sell workflows.