Buyers and their counsel use these words every day. Most owners meet them for the first time in a contract with a deadline on it. Read them here first, with no deadline.
Processing records so that individuals cannot reasonably be identified, with no key kept to reverse it. Stronger than pseudonymization, and rarely perfect: no process removes re-identification risk entirely, which is why a license names the method rather than promising an outcome.
A license clause letting the seller (or an independent reviewer) confirm the buyer did what the contract says: de-identified before training, used the records only within the field of use, and deleted the supplied copy on schedule.
A buyer's firm terms, price included, that become a contract when the seller accepts. Follows sample review and diligence. Distinct from the indicative range that arrives first.
A dataset with a similar shape to yours (industry, years, systems, modalities, volume) that has already been priced or licensed. Comparables are one of the three inputs to pre-market pricing.
Putting a qualified manifest in front of several fitting buyers at once, with the same information and the same deadline, so offers can be compared on price and terms. The alternative is taking the first number from the first lab that called.
A company that builds training and evaluation datasets for AI labs and model builders, often under contract to several at once. Data labs are frequent buyers of single-company operational records.
In privacy law, a party that handles personal data on someone else's behalf. A seller-side agent that never receives the records is not a processor of them; the buyer, once records are delivered, is.
A controlled space where a buyer reviews samples or documents under NDA during diligence. In this process the data room, if one is used, belongs to the seller or the buyer, never to the agent.
Removing or replacing details that identify people or companies in a record (names, emails, phone numbers, account numbers) before the record is used for training. Done by redaction, transformation, or both, to a method named in the license.
A written confirmation from the buyer, signed by a named person, that the supplied copy of your records has been destroyed at the end of the license term. Deletion does not reverse training already completed; the field-of-use and no-re-licensing clauses cover that.
What vetted buyers are currently asking for and what they are paying for it: industries, modalities, years, volumes. Maintained continuously so a new manifest can be benchmarked in hours rather than shopped for weeks.
The buyer's review before binding terms: checking the sample matches the manifest, confirming the seller's rights to license, and asking questions about provenance and gaps. Usually one to three weeks.
Delivery of the licensed records from the seller's systems to the buyer's, after signing, with no intermediary copy. The structural reason a seller-side agent cannot misuse your data: it never has it.
The two-page agreement between a seller and the agent. States that the engagement is non-exclusive, that no upfront fee applies, the success-fee schedule, how either side can end it, and the tail clause.
An internal group at a large non-AI company building agents or assistants for its own operations. Increasingly a buyer of operational data from businesses in adjacent or supplier industries.
Records held back from training and used to test whether a model handles real cases correctly. Sets with clear outcome labels (resolved, won, paid) are especially valuable as evaluations, sometimes more than as training data.
A buyer's right to be the only licensee of a dataset, for a period or a field of use. Worth paying more for from the buyer's view, worth charging more for from the seller's. Always a separate, priced term; never assumed.
An offer that expires in days, designed to prevent comparison. Common enough in this market that it appears on every red-flags list, including ours. Offers we pass along carry at least ten business days to decide.
The license clause defining what the buyer may do with the records: train a model, evaluate one, build a specific product category, or all three. Narrower fields of use are worth less to that buyer and leave more value for others.
A company training general-purpose foundation models at the largest scale. Frontier labs buy directly and through data labs, and are among the most demanding on rights and provenance.
The first number you see: a low and high figure benchmarked against current buyer demand and comparable sets, with the fitting buyers' confirmation. Arrives within 24 to 48 hours of a complete manifest. Not a quote; binding terms follow sample review.
A service that passes a seller's details into one buyer's referral program and is paid by that buyer when a deal closes. Free to the seller, which is the point: the router's incentive is to close with the buyer that pays it, not the buyer that pays you.
How long the buyer may hold and use the supplied copy. Commonly 90 days for delivery-and-train arrangements, longer for evaluation sets. Deletion obligations start when the term ends.
A one-page, metadata-only inventory of a company's records: which systems, which record types, how many years, approximate volume, and how the records connect into workflows. Contains no records and no files. The document buyers price from.
A self-serve platform where sellers list datasets and buyers browse. Charges listing fees, commissions that can reach 30%, or both, and leaves pricing and negotiation to the seller. The opposite of representation.
Information about records rather than the records themselves: counts, date ranges, system names, field lists. A manifest is metadata. Pricing and the first offer happen on metadata alone.
The form a record takes: text (email, tickets, documents), structured data (CRM fields, invoices), audio (calls), video (recorded meetings), or code. Buyers often have gaps in one modality, which is why call recordings or chat logs can lift a set's value disproportionately.
A company fine-tuning or training models for a specific vertical or task, such as accounting agents or support automation. Smaller than frontier labs, often more specific about the industries they need.
A mutual confidentiality agreement signed between you and a buyer before any sample moves. The manifest needs no NDA because it contains nothing confidential; samples always do.
A payment from a buyer to an intermediary for sourcing a dataset. Legitimate when disclosed; a conflict when hidden. We do not take origination fees on sellers' deals, so there is nothing to disclose.
A field in a record that says how it ended: ticket resolved or escalated, deal won or lost, invoice paid or disputed. Records with outcome labels teach models what works, and buyers pay more for them.
Any detail that identifies a person: names, emails, phone numbers, addresses, account or card numbers, and sometimes combinations of less obvious fields. The target of de-identification.
Where records came from and how they were produced: which company, which systems, which years, generated by whom in the course of what work. Buyers evaluate provenance first because it determines whether the data is real, licensable, and representative.
Replacing identifiers with consistent stand-ins (Jordan Lee becomes Person 014 everywhere) so the workflow still reads correctly without exposing the real person. Reversible if a key is kept, which is why the license says who keeps it and for how long.
A buyer passing your records on to a third party. Prohibited in every license we introduce. One of the five terms a buyer accepts before seeing a manifest.
Removing a detail outright rather than replacing it. Used for anything that cannot be safely transformed: card numbers, health details, privileged passages.
A buyer's option to license newer records on the same terms after the first delivery, usually for an additional payment. Turns a one-time license into recurring revenue if your records keep accumulating.
A post-training method in which human judgments steer a model toward better responses. Records that show how experienced staff resolved a case, and whether the customer was satisfied, are a natural source of that judgment.
An intermediary that represents the company licensing the data, is paid only by that company, and never takes possession of the records. Prices the data, runs the competitive process, and negotiates terms. Not a marketplace, not a buyer, not a data processor.
Training a model on examples of a task done well. Operational records (a ticket and its resolution, a brief and the delivered work) are exactly that kind of example, which is why they are licensed.
A fee paid only when a license closes and the seller is paid. Ours is tiered and published: 12% of the first $250,000 paid to you, 8% of the next $750,000, 5% above $1,000,000. Nothing closes, nothing is owed.
The copy of your records delivered to the buyer under the license. Your originals stay in your systems. The supplied copy is what the license term, deletion obligation, and audit right apply to.
The one protection in a non-exclusive engagement letter: if you close with a buyer the agent introduced to you by name within 12 months, the success fee still applies. Does not cover buyers you found yourself.
A sequence of steps across systems that gets a piece of work done: an email becomes a ticket, gets approved, gets resolved, gets invoiced. Models learning to act, not just answer, are trained on trajectories, which is why connected systems are worth more than isolated ones.
How much of the real work leaves a trace in a system. A team that logs every call, ticket, approval, and handoff has dense records; a team that runs on hallway conversations does not. One of the strongest predictors of value after years of history.
Definitions describe how these terms are used in this market and in the agreements we see. They are not legal definitions and not legal advice. Your counsel's reading of a specific contract controls.
Seven questions. A one-page manifest. A first offer in 24 to 48 hours.
Measured from a complete data manifest. A first offer is a process target, not a guarantee; buyers decide eligibility, price, and terms.