Can I get data from a company that is shutting down?

Yes, and they are cheaper: reported at $10,000 to $100,000 per archive through wind-down firms. The trade-off is that the data stopped growing and the staff who could answer questions are gone.

How fast can a deal close?

Aggregators report about two weeks from a seller's upload to payment once terms are agreed. The rights check before that is what varies: days for a company with simple contracts, longer for an agency or a firm with client-held material.

Do you handle non-English data?

Rarely today. De-identification tooling is English-first and buyers say the same. Tell us the language anyway; demand changes.

Where to buy licensed company workflow data for training

The four routes a lab or applied AI team can take to get operational data from working companies, what each one costs in time and money, and how to judge provenance.

There are four ways to get operational data from working companies: buy direct from a company, buy from an aggregator that has already licensed and de-identified data, license through a marketplace, or work through a broker who brings you competing sellers. Direct deals take quarters and only happen with brand-name catalogs. The other three take weeks. Which one fits depends on whether you need a specific shape of data or a volume of a general type.

Route one: direct from the company

This is how the headline deals happened: News Corp to OpenAI, Reddit to Google, Spirit Airlines to Google in bankruptcy court. Every one of those sellers had a legal team and a data team. A 60-person logistics company does not, which is why a lab's procurement process rarely reaches one. If you want a specific company's data, expect to do the rights work yourself, pay for de-identification, and wait.

Route two: aggregators that buy as principal

micro1, Polyshares, Grepped and others license data from companies, pay the seller themselves, run de-identification, and resell packages or build environments on top. You get one counterparty, a catalog, and a package that already carries a provenance memo. You pay their margin, which is not published, and you take what they have sourced. micro1 has committed $1 billion over twelve months to this. Grepped's founder reports a typical deal of $200,000 to $300,000 and about two weeks from upload to payment on the seller side.

Route three: marketplaces

Troveo and similar platforms sign one license with a seller and resell non-exclusively to many buyers. Good for volume across a general type (support tickets, chat) and for media. Less good when you need a particular industry's workflow or exclusivity, and the same corpus may already be in a competitor's training set.

Route four: a broker on the seller's side

We represent the company that owns the data: check what it can legally license, write a one-page summary, and put it in front of more than one buyer. For you that means:

Data you specify by industry, system and outcome fields, not whatever is in a catalog.
A provenance memo before the summary: ownership confirmed, client-held material excluded, customer-contract clauses on de-identified use.
De-identification by a third-party vendor, with the seller reviewing a sample before acceptance.
Exclusive or non-exclusive, with a term. Two years is what sellers accept. Sellers are steered away from perpetual terms, so plan on a term or a renewal.
A finder's fee from the buyer on close, agreed before any introduction.

How to judge provenance, whichever route

Ask who owned the data on the day they were made. If the answer involves "on behalf of a client," stop.
Ask to see the de-identified-use clause in the seller's customer agreements, or the memo that summarises them.
Ask who did the de-identification and how pseudonyms are kept consistent across files. Redaction that breaks links is a sign of home-made work.
Ask what was excluded and why. A package with nothing excluded was not reviewed.
Ask for the seller's sample review sign-off. It is your protection as much as theirs.

What to put in a request

Industry, record types, the outcome fields you need (resolution codes, won or lost, approval states), minimum years of history, language, exclusivity preference, and whether the use is training, evaluation or environment construction. The more specific the brief, the faster the answer on what exists.