Discuss a Dataset

Our approach

Careful by design, practical in delivery.

How we take data from an initial conversation to a licensed dataset. What follows describes our approach. The exact steps for any dataset depend on the data, the jurisdictions involved and the intended use.

The path from source data to licensed delivery

  1. Source systemsData held by the partner
  2. Rights reviewAuthority and permitted use
  3. Scoped exportOnly what the use requires
  4. StructureConsistent, documented formats
  5. Quality and privacyChecks and de-identification
  6. Release reviewPartner sign-off before licence
  7. Licensed deliveryDefined uses, secure transfer
Review gates involve people and the data partner. They are never automated away.

Each part of our approach

Data sourcing

Our model is built on direct relationships with organisations that hold data, not on web scraping. We prioritise data types that AI developers are actively asking for, and we tell potential partners early when we don't see a fit.

For each engagement

  • Start from a description of systems, data types and history
  • Compare against current buyer requirements before going further
  • Agree confidentiality terms before any detailed review

Rights and authorisation

Possessing data is not the same as having the right to license it. We work with each data partner and their advisers to establish what can be done, and what can't.

Questions we work through

  • Who controls the data and its copyright, and under which contracts and platform terms
  • What authors, contributors, employees, customers and other parties agreed to or were told
  • Whether there is a lawful basis for any new use of personal data
  • Which regulatory, contractual or confidentiality restrictions apply
  • Which uses, if any, can be permitted under a licence

Extraction and transformation

Once a dataset is approved, data is exported through methods the partner agrees to, ideally with the partner running the export. It is then converted into consistent, documented structures. We select tooling for each engagement rather than relying on a single platform.

Depending on the dataset

  • Partner-controlled exports, scoped as narrowly as the use allows
  • Scope filters applied as early in the process as possible
  • Normalised formats with documented schemas
  • Conversation and thread structure preserved where it matters

Data quality

Useful data is coherent, representative of real work and free of noise that would mislead a model. Quality checks are designed around the buyer's stated use.

Typical checks

  • Completeness and continuity of threads and records
  • Removal of duplicates, automated messages and system noise
  • Sampling-based human review
  • Documented known limitations, rather than silent gaps

De-identification and privacy

Removing names is not the same as anonymisation. De-identification is designed for each dataset, and residual risk is assessed rather than assumed away. De-identified data can still be personal data, and the obligations that come with it may continue to apply.

Depending on the dataset

  • Direct identifiers removed or replaced with consistent pseudonyms
  • Indirect identifiers considered, such as job titles, locations and rare events
  • Free text reviewed for embedded personal or confidential details
  • Licence terms that prohibit re-identification attempts

Provenance documentation

A dataset is only as trustworthy as the record of where it came from. Each licensed dataset is intended to travel with documentation that a buyer's legal and research teams can rely on.

What the documentation covers

  • Source description, collection period and coverage
  • Summary of authorisation and agreed uses
  • Processing and de-identification steps applied
  • Residual risks and known limitations

Licensing and delivery

Licences are specific to a dataset and a purpose. They set out what the buyer may do with the data, what they may not, and what happens when the licence ends.

Licence terms typically address

  • Permitted and prohibited uses
  • Redistribution, exclusivity, term and territory
  • Security obligations and deletion or return at the end of the term
  • Delivery through an agreed secure transfer method, with access limited to named parties

Ongoing partner relationships

Some partners generate relevant data continuously. Where that suits both sides, a one-off licence can become an ongoing supply arrangement, reviewed as needs change.

What that can involve

  • Periodic refreshes under the same rights framework
  • Regular review of terms, uses and controls
  • Clarity on how future supply can be paused or ended, subject to contract

Questions about how this would work for you?

Every dataset is different. Tell us about yours, or about what you need, and we'll walk through how the approach would apply.

Become a Data PartnerDiscuss a Dataset