A data deal lets a buyer use a company’s information for an agreed purpose. The company may earn revenue from material it already creates, while the buyer gains examples for developing or testing AI. Getting to that agreement requires people who understand the data, people who can approve its use and a buyer who can explain what it needs.
SnowRock helps find potential matches, arrange introductions and coordinate the conversation. The company and buyer make the licensing decisions. The sequence below shows how to move from interest to delivery without leaving the important work until the end. Some steps can run together, but each should end with a clear decision.
1. Describe what the company has
The company starts with a short inventory. Name the kinds of information, where they are stored, how far back they go and who uses them. A software business might have support conversations linked to bug reports. A distributor might have orders linked to shipping changes and returns. A professional services firm might have project briefs, reviewed drafts and final work.
Ask someone who works with those systems to check the description. An estimated file count can help plan a conversation, but it should be labeled as an estimate. Explain gaps, such as missing attachments or older files stored in a different system. SnowRock uses this description to judge whether a buyer conversation makes sense, rather than asking the company to export everything first.
2. Define what the buyer needs
The buyer describes the work its AI should learn or the questions it wants to test. A request for customer service data is too broad to guide a search. A request for conversations about delivery problems, linked to the action taken and the final outcome, gives the company something specific to assess.
The brief should distinguish training, which changes how a model responds, from testing, which measures how well it performs. It should also name required languages, time periods, file formats and exclusions. The buyer appoints a reviewer who can make the sample decision. SnowRock compares this request with the company’s description and brings the unresolved questions into the first discussion.
3. Approve an introduction and the information to share
Before an introduction, establish which company details can be shared with which buyer. The owner may be comfortable sharing a description of its support workflow but want to withhold customer names, examples and commercial details until confidentiality terms are in place. SnowRock coordinates that boundary and confirms who should join the conversation.
The company identifies the person who can approve a license and the staff who can check restrictions. A shared drive can contain employee work, customer material and third-party content with different permissions. The U.S. Copyright Office explains that creators, employers and parties receiving transferred rights may hold copyright. Access to a file alone does not settle who may license it. Resolve relevant ownership, contract and privacy questions before sharing a sample.
4. Review a small, approved sample
The buyer and company agree the sample’s purpose, contents, permitted access and treatment after review. They also decide whether preparing it is paid work. The sample should answer the difficult questions: can the buyer open the files, follow the sequence of events and find the information needed for its task?
Include ordinary cases and the situations that could change the purchase decision. In a returns workflow, that might mean a completed refund, a rejected request and a case with no recorded outcome. These are illustrative choices, not a universal sampling rule. The buyer should explain why its chosen mix is enough for an initial decision.
Google’s data-quality guidance identifies missing values, duplicates and incorrect labels as common problems. Record those findings in the review. If the sample fails, decide whether a targeted correction is worth testing. Do not expand the order while the basic fit remains unclear.
5. Agree the license and commercial terms
Once the sample shows a plausible fit, the parties can negotiate a defined package. The price should refer to that package and its permitted uses. It should be clear whether the agreement includes one delivery, later updates, preparation work or ongoing support. An attractive headline payment can be less attractive if the company must fund substantial extraction and review.
The buyer pays the company for the license. SnowRock’s buyer-paid introduction or sourcing fee is agreed separately. A sourcing fee pays for the specified search work; a license payment pays for agreed use of the data. The agreement should say what happens if preparation costs rise, fewer examples qualify or the buyer requests a wider use.
| Decision | What to settle |
|---|---|
| Use | The allowed AI tasks, users and any onward sharing. |
| Scope | The included data, excluded material and delivery period. |
| Duration | How long access or use continues and what happens when it ends. |
| Exclusivity | Whether the company may license the same material elsewhere. |
| Acceptance | The checks, review window and remedy for a failed delivery. |
| Payment | The amount, currency, invoice requirements and due dates. |
6. Assign preparation and transfer work
Name the person or provider responsible for each delivery task. That may include exporting files, connecting related records, removing excluded material and producing a field dictionary that explains each column. Agree who checks the result and who pays for rework. SnowRock’s coordination does not replace the technical work needed to prepare a dataset.
The FTC’s business guidance recommends limiting sensitive information and access to what is needed, protecting transfers and assessing service providers. Apply those decisions to this delivery: use an agreed transfer method, name the recipients and set access limits. Keep a versioned copy of what was delivered so later questions refer to the same files. Avoid using an ordinary outreach email thread to exchange sensitive exports.
7. Check the delivery against the agreement
The buyer’s reviewer checks the package within the agreed review period. Start with the promised files, formats and connections, then inspect the content against the sample criteria. Receiving a download link is not the same as accepting the data. Likewise, a disappointing experiment does not automatically mean a delivery failed its agreed requirements.
If something is wrong, identify the affected files and the unmet requirement. The company or preparation provider can then correct, replace or explain the issue under the agreed process. SnowRock helps keep that discussion specific. Separate defects from new requests: adding a language, changing the task or demanding a new annotation scheme may require a revised scope and price.
8. Complete payment on the agreed schedule
Payment follows the contract’s milestones. Those might include an advance, an accepted sample, a completed delivery or another clearly defined event. There is no single payment schedule for every data deal. Both sides should know which event makes an invoice payable and which person approves it.
The company sends the required invoice and confirms payment details through its normal finance process. The buyer pays the license amount; SnowRock invoices its own agreed fee separately. If payment is late, the parties use the notice and resolution steps they agreed. SnowRock cannot guarantee the buyer’s payment or replace the company’s decision about acceptable commercial risk.
9. Keep track of obligations after delivery
Keep a named contact on each side for corrections, permitted-use questions and any promised updates. Record the version, delivery date and relevant terms together. Hugging Face’s dataset-card guidance describes documenting a dataset’s contents, context and license information; a short private delivery note can serve a similar practical purpose.
If an error is discovered later, explain what changed and which files are affected. If the buyer wants another use, return to the license terms before proceeding. Discuss retention, deletion and the treatment of outputs or trained models when negotiating the deal. Deleting a source file and reversing its effect on a trained model are different questions that should not be left implicit.
Start with the description, then decide the next step
Consider a fictional retailer with support tickets linked to returns. The owner describes the systems and available history. A buyer asks for examples of difficult return decisions. They approve a small sample, discover that older outcomes cannot be linked reliably and narrow the proposed delivery to a later period. That change improves the agreement before either side commits to the full preparation cost.
The same sequence can work for software issues, financial review files or project revisions. The first action is simple: tell SnowRock what data your company has, or describe the data your team needs. Include the person who understands the files. The next step is a focused conversation about fit, permissions and whether a sample is worth preparing.