What an AI data supplier should tell you before you ask

A supplier with a real process has the artifacts on hand. The disclosure set that should arrive unprompted, and how to request each item when it does not.

Volunteered information is the first signal

The most useful thing to observe in a data procurement is not what a supplier can answer but what they raise on their own. A supplier who recorded the material or managed the recording has the documents at hand, because they needed them to run the project. A reseller has to go and ask someone, and the delay is visible.

The items below are the disclosure set a serious supplier hands over as a matter of course. None of them are trade secrets. All of them are things a buyer needs in order to decide, and their absence is information in itself.

The rights chain, from the session to your license

This is the part buyers most often accept as a sentence rather than a document. The sentence is usually true and almost never sufficient.

  • Who recorded the material, in which country, and in which year. Aggregated corpora often cannot answer this per file, which is itself the answer.
  • The consent text itself, not a summary of it. A general media release usually does not cover AI training, and the difference shows up only in the wording.
  • The chain from the recording to the license being offered: who holds what, and where the right to sublicense to you was granted.
  • The distinction between owning the data and having the right to license it for training. Those are different statements, and a supplier who treats them as one has not examined their own position.
  • Whether the material has been licensed before, and on what terms. A corpus that has circulated for years may already sit inside public models, which changes what it can be used to measure.

What is actually in the corpus

Composition disclosure is where aggregated catalogs go vague, and the vagueness has a pattern. Ask for the speaker manifest and for the number of separate recording efforts the corpus draws from.

  • Distinct speakers, and the share of total hours held by the largest contributors. A corpus where a handful of voices dominate will overfit no matter how large the total is.
  • How many separate recording programs the material comes from. A catalog assembled from several sources inherits several consent standards, several recording setups, and several annotation conventions at once.
  • What was removed before delivery, and why. A supplier who cannot say what was discarded cannot describe the floor of the dataset being sold.
  • Whether any listing overlaps with another listing from the same supplier. The same underlying collection re-described under a second title is common, and the overlap becomes visible only when you compare sample files.

The annotation conventions and the quality method

Annotation disclosure is a document request rather than a conversation. If the answer arrives as a description, ask for the document behind it.

  • The guideline with its version and date, including the convention decisions that matter: filler words, numerals, casing, unintelligible stretches, and how a mid-sentence correction is written.
  • Whether any subset was annotated twice, and what the measured agreement was. No figure means consistency was never measured, which is not the same as consistency being fine.
  • Who annotated, and what training they received beyond fluency in the language.
  • The quality report from the most recent comparable delivery: what was sampled, what failed, and what was done about the failures.

How they handle the questions you have not thought of

The last group is about process under strain, and it is where rehearsed answers run out.

  • What happens when a speaker withdraws consent after the material has shipped.
  • Who is accountable when a delivery is rejected, and what the last rejected delivery cost in schedule terms.
  • What a specification change mid-project costs, and who prices it.
  • What the supplier would need to check before answering a question about one specific file. A supplier with a working record system can name the lookup; one without will offer a call instead.

How to ask when it does not arrive

Requests work better than questions, and written requests work better than calls. Four habits cover most of the gap. Keep one expectation in mind throughout: a supplier who cannot produce the consent text and the guideline is usually not withholding them. They do not have them.

  • Ask for the artifact rather than an assurance. A request to send the consent form template can be answered only by producing something, while a question about whether consent is obtained can be answered with one word.
  • Ask in writing and keep the written answer. Spoken claims do not survive a change of account manager, and the written version is the one that will describe what actually happens.
  • Set a date for the attachments and note what arrives by it. What arrives late and what never arrives are two different findings.
  • Treat a substitute as an answer. An offer to walk through the documentation on a call, in place of sending it, is a response about the supplier rather than about the documentation.

More insights

Submit a sourcing request

Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline — not a range. If we cannot source it well, we will tell you that instead.

  • Pilot batch before the full run, so problems surface early.
  • Consent documentation delivered with the data.
  • No medical or clinical data. No recorded telephone calls.

We reply within two business days. Your details are used only to answer this request. See our privacy policy.

Contact

Talk to a human

Send a specification and we will come back with a real number and timeline.

Submit a sourcing request

Or email hello@linguacorpus.com