How AI labs review a data sample before licensing

AI labs review a data sample by first confirming it is de-identified, then verifying it shows a complete, traceable workflow, and finally scoring its relevance, volume, and uniqueness against their training goals. The review follows a three-phase process that we manage for you as your broker.

Why labs need a sample

A data sample is the first concrete proof that the operational records you call "business records" can be used for model training or reinforcement-learning (RL) environments. Labs cannot base a deal on a questionnaire alone; they must see the actual shape of the data to judge quality, continuity, and visible reasoning. A well-chosen sample also demonstrates that you can deliver the larger set under the same conditions, which lowers risk for the buyer.

What labs look for in a sample

When the sample lands in the lab's sandbox, the data team runs a concise checklist. The key items are:

1. De-identification. All personally identifiable information and any company-specific confidential tags are removed or replaced with stable pseudonyms such as "employee 632". The lab never sees raw files.
2. Operational chain. Each record must show a clear start-to-finish path, for example a support ticket that includes the initial request, the actions taken, and the final resolution.
3. Continuity. The sample should span multiple years on the same tools, proving consistent formatting and metadata.
4. Reasoning visibility. Notes, chat excerpts, or decision logs that explain why a particular action was taken are highly valued.
5. Industry rarity. Data from a niche sector or a workflow that few other sellers provide can raise interest and price.

If any of these elements are missing, the lab asks for a revised sample before moving forward.

Need licensed data for a model? Tell us what you need

How the review proceeds

The lab's review mirrors the safeguards we put in place as your broker.

1. Initial intake. We forward a one-page summary of your answers together with a small, de-identified excerpt. No data leaves our secure workspace until you sign off on the quote.
2. Technical assessment. Engineers load the sample into a sandbox, run scripts that check schema consistency, missing fields, and duplicate entries, and then perform a manual spot-check to confirm that the captured reasoning is understandable.
3. Fit-score report. After the technical pass, the lab issues a short report that grades the sample on relevance, volume, and uniqueness. That report becomes the basis of the offer they will make to you.

Because the lab never sees raw PII, the de-identification step is critical. In some deals the buyer handles de-identification; in others the seller does it first to improve the offer. The lab's contract will specify who performs the cleaning and who pays for it.

Common pitfalls for sellers

Even a solid sample can hit snags that slow the lab's review. The most frequent issues are:

• Over-cleaned data. Stripping away too much context, such as the decision rationale, leaves a flat log that loses value.
• Inconsistent formats. Mixing CSV, JSON, and PDF files without a clear mapping forces the lab to spend extra time harmonizing the data.
• Missing metadata. Dates, department tags, or tool identifiers are essential for the lab to understand the timeline and scope of each workflow.

A quick internal test can catch these problems early. Pull five random records from your source system, de-identify them using our service or your own pipeline, and ask a colleague who wasn't involved in creating the data to read them. If that colleague can reconstruct the full business process, you are likely ready for the lab's review.

What you do next

If you have identified the operational data you want to license, tell us exactly what you need. Use the button below to submit a brief request describing the type of records, the expected volume, and any special workflow elements you consider valuable. We will match you with a buyer, arrange the secure sample exchange, and keep you informed at each stage. For more detail on how a buyer request runs, see our buyer request page. You can also read our guide on licensing your company data to AI labs for additional context: how to license your company data to AI labs.