For AI labs and data buyers
Operating records from real US firms, licensed for AI training
Pewter licenses de-identified copies of business records that were never on the public web: the email, tickets, documents and system records that show how a job went from the first request to the result. Request the catalog to see what is available, or tell us what you need next.
What we can bring
Real work, recorded step by step
Operating records, not public text
Email, tickets, documents, workpapers and system records showing how quotes get priced, claims get resolved and projects get delivered.
Owners who have said yes
Every firm approves its terms in writing, with its exclusions, before anything is delivered.
De-identified before delivery
People, contact details, account numbers and credentials are removed or replaced with consistent placeholders, so the work stays intact.
Written terms
Scope, permitted use, exclusivity and payment are set out in a license agreement. How we handle records.
For researchers
What these records are useful for
Long-horizon tasks
A long-horizon task takes a person hours or days and many steps that depend on each other. A freight load from first quote to paid invoice is one, and so is an engineering job from proposal to closeout. We look for corpuses where that whole chain can be followed.
Trajectories with an outcome
A trajectory is the recorded path from the start of a piece of work to its result. Records that show how the work ended, such as paid, denied, approved or reworked, give a verifier something to check an attempt against.
Material for RL environments
An environment needs the realistic content of a company at work: its messages, documents and system records. Each catalog profile says which systems a corpus covers and how its records link across them.
Unpublished records for evaluation
These are internal records their owners have not published. We ask every owner about earlier buyers and licenses and report what they tell us, so you know what a held-out set is built on.
AI training data
What is AI training data, and where do labs get it?
AI training data is the material a model learns from: text, records, images, audio or recorded actions. Labs get it from five kinds of source, and each teaches a model something different. Pewter supplies the fifth: licensed operating records from real firms.
| Source | What it is | What it shows a model |
|---|---|---|
| Public web text | Pages, forums and code collected from the open web. | Language and general knowledge. |
| Licensed published content | Books, news and reference works licensed from their publishers. | Edited, finished writing. |
| Commissioned work | Tasks written or graded by hired experts and annotators. | How an expert answers a set question. |
| Synthetic data | Examples generated by other models or by simulation. | More of a pattern that is already known. |
| Licensed operating records | The email, tickets, documents and system records a firm kept while doing its work. | How a real job went from the first request to the result, including what went wrong. |
Operating records were never published, so they are not in a web crawl. They are also the only one of the five that records work as it happened, not as someone later described it. That is why they suit long-horizon tasks and evaluation.
What we represent
Training data by sector
Freight and logistics
Rate quotes, dispatch records, shipment documents, delivery exceptions and claims.
Accounting
Workpapers, reconciliations, financial models and tax-preparation files.
Engineering and construction
Drawings, calculations, technical reports, revisions and project specifications.
Distribution
Order history, supplier records, product knowledge and customer correspondence.
Trades and home services
Estimates, job records, schedules and the messages around them.
Professional services
Client correspondence, project files and the working documents behind a finished piece of work.
What is available changes as owners sign. The catalog shows the current corpuses, each with its record types, years of history and known limitations. New to the terms? Read what AI data licensing is and how a deal works.
The corpus catalog
See what is available before you write a brief
The catalog holds anonymized profiles of the corpuses in our pipeline. Each profile gives the industry, record types, years of history, volume, how the records link across systems, readiness and known limitations. It can be filtered by industry, record type, depth and the model capability the records may support. It is private and opens by link, so we send one to each buyer.
Send a request
Tell us what data you need next
Sector, workflow, systems, years of history and volume all help. We reply within one business day. Prefer email? Write to lawrence@pewterdata.com.
Preparing your brief? Read how to write a data request and how to evaluate a sample.
Before you ask
What buyers ask first
- What is AI training data?
- AI training data is the material a model learns from: text, records, images, audio or recorded actions. Labs draw on public web text, licensed published content, work commissioned from experts, synthetic data, and licensed operating records from real firms. Pewter supplies the last kind.
- Where can an AI lab buy training data from real companies?
- From the company itself, from a firm that resells data it has licensed, or through Pewter Data, which licenses de-identified operating records from established US firms and sells them to AI labs directly. Send a request with the form on this page, or ask for the catalog.
- Is Pewter Data an AI data marketplace?
- No. A marketplace lists datasets for anyone to browse and buy. Pewter's catalog is private and opens by link, each license is agreed with the owner in writing, and buyers review a sample before any full export.
- What does Pewter Data license to AI labs?
- De-identified copies of operating records from established US firms: email, tickets, documents, workpapers and system records. The records show how work was done from the first request to the result, in sectors such as freight and logistics, accounting, engineering and construction, distribution, trades and professional services.
- Does Pewter have a catalog of available corpuses?
- Yes. The catalog is private and opens by link. It holds anonymized profiles of the corpuses in our pipeline, each with its industry, record types, years of history, volume, readiness and known limitations. Request it with the form on this page and we reply within one business day.
- Can the records support long-horizon tasks and RL environments?
- That depends on the corpus, which is why each profile says how its records link across systems. Corpuses where a unit of work can be followed from the first request to a recorded outcome are the ones suited to long-horizon tasks, and they supply the realistic company material that an RL environment is built around.
- Have these records been published or licensed before?
- They are internal operating records that their owners have not published. We ask every owner about earlier buyers, brokers and licenses, and we report what they tell us.
- How are the records de-identified?
- People, contact details, account numbers and credentials are removed or replaced with consistent placeholders before delivery, so the same person or company stays linked across a workflow without being named.
- Can a buyer review a sample first?
- Yes. Buyers review a representative sample before any full export begins.
- Are licenses exclusive?
- Exclusivity is set deal by deal. Scope, permitted use, exclusivity and payment are written into the license agreement.
More detail: how AI labs buy training data and how we handle records.