Data forthe latestfrontier modelsfrom Anthropic, OpenAI, Google, Meta, Mistral and xAI.
You already hold years of emails, tickets, orders and records. AI labs pay to train on data like that. We take out everything that identifies a person, license what remains to the labs, and pay you a share of every fee.
What we do, in plain terms.
The companies building the latest AI models pay to train on real business records. Most businesses have years of them and have never been paid for one. We turn that into a licence, and we handle every step between your export and the lab.
-
01
You already have the data
Support email, chat logs, orders, invoices, CRM notes, documents. Every business produces them, and once the job is done they sit in a system doing nothing.
You do: tell us which systems you run. We say which records qualify, and which never leave.
-
02
We take the people out
Names, email addresses, phone numbers and account numbers are suppressed or replaced before a copy leaves. A lab receives how your business runs, never who your customers are.
You keep: the data itself. We hold a copy under a licence from you, and the exclusion list you write goes into the agreement.
-
03
Labs license it. You get paid.
We sign the licence with the lab in our own name and carry the warranties. You hold a named share of every fee your data earns, fixed in writing before a single file moves, renewals included.
You get: a statement each quarter showing what was licensed and what you are owed, with an audit right to check it.
It starts with the systems you already run. Operating records from these, licensed to the labs that need real records rather than synthetic ones.
Whatever identifies a person comes out before delivery
One record, three columns, nothing hidden. On the left, the field as it sits in your system. In the middle, what the published standard does to it. On the right, what a lab receives. Switch the source to see your own kind of business.
Source field
The field as the source system names it.
Technique applied
The treatment the standard sets for that class: suppressed, replaced with a token, generalised, scanned, or none required.
As licensed
What a lab receives: a bar where the value was removed, or the token, the generalised value or the kept value.
The technique column is drawn verbatim from the published standard, and no technique shown here can be reversed.
Labs see exactly what they are buying, before they buy it.
A manifest lists which systems a corpus comes from, how much there is, which years it covers, how many people contributed and what was removed. A licence is negotiated against it. Below is a redacted sample from an anonymised supplier, each column keyed to a note beneath it.
| Source system | Volume | Years | Contributors | De-identification |
|---|---|---|---|---|
| Mail, tenant wide | 30 mailboxes | 2022 to 2026 | 27 | Addresses tokenised, bodies scanned |
| Chat, 46 channels | 7,900 messages | 2022 to 2026 | 27 | Handles tokenised, DMs excluded |
| Application database | 140 tables | 2021 to 2026 | 70 | Identifier columns suppressed at source |
| CRM | 5,539 accounts | 2022 to 2026 | 12 | Buyer identity dropped |
| Accounting | 5 fiscal years | 2021 to 2026 | 4 | Bank and routing fields suppressed |
| Document store | 3,769 native files | 2021 to 2026 | 70 | Scanned pages redacted by pixel destruction |
The system of record it was exported from, named specifically.
Counted from the source system at export, in the unit that system uses.
The continuous period. Any gap is stated, never smoothed over.
Distinct people whose work appears in the system.
The technique applied to that source before anything leaves our systems.
Who holds the data, at every step.
We take custody, and every step names who holds the data and under which agreement. Nothing is uploaded until an agreement exists.
Chain of custody
Each transfer is an event with a holder and an instrument. The raw copy is destroyed on a fixed schedule and the destruction is recorded.
- Export received under Supplier agreement
- Rights record checked against Schedule B
- De-identified under Standard v1.0
- Substring scan run, report stored 0 matches
- Licensed to the lab as principal Licence
- Raw copy destroyed and certified Certificate
What never leaves your building
Suppliers can add to the standing exclusion list and we hold them to it. Nobody removes from it. All six are Schedule B exclusions in the supplier terms.
- Customer records and anything identifying a customer
- Payment, credential, biometric and precise-location data
- Private messages, HR and grievance channels
- Anything under a third-party confidentiality agreement
- Data about minors, and student records
- Anything the company did not itself collect
What a lab is actually buying
Provenance on every file, coherence across systems, exact sizing, native formats kept, deduplicated, segmented by industry.
The licence, in summary
Field of use priced and granted separately, a defined term, non-exclusive by default, an indemnity that is ours and capped. Executed as principal.
For the companies that hold the data
Three protective questions, answered before the commercial one.
What actually leaves your building
Business-system records only. Never customer data, never payment or credential data, never private messages or human-resources channels. You write the exclusion list and it goes in the agreement.
Who holds it afterwards
We do, under a licence from you, for as long as the agreement runs. Raw identifiable copies are destroyed on a fixed schedule and the destruction is recorded.
What you are exposed to
You warrant that you own or hold the rights to what you give us, and that giving it to us does not breach an agreement you are already in. Where employment or privacy law applies, we tell you the steps and give you the templates.
What it pays
A named share of every licence your data earns, fixed in the agreement before a single file moves. It applies to renewals and reuse.
Posture, with its sources named.
Every claim on the trust page names the internal document it comes from, and we will show you that document under NDA.
Operating data, licensed direct.
Tell us which side of the table you are on and we will send the relevant paperwork first.