Business data for AI agents

AI dataset delivery and acceptance checklist

Last updated November 7, 2026/Business data for AI agents

An AI dataset delivery acceptance checklist is for the team that receives bytes, not for the team that liked a profile. They verify that the format is one they can read, that the schema matches the brief, that documentation says what the fields mean, that the transfer used the controls you agreed, that the version is the one you reviewed, and that integrity checks match what the sender claimed. They log discrepancies and they do not call the set accepted until those lines pass. Arrival in a bucket is not acceptance.

This list is a buyer practice. It is not a description of a delivery platform Big Data/Little Guy operates, and it is not a promise that every transfer we might someday coordinate will include every control below. Where a handoff exists, ask which of these checks the parties will actually perform. If a control is not offered, do not write it into your runbook as if it were. The sourcing path that leads here is how a buyer sources proprietary business data. The sample review that should precede a full transfer is how to diligence a business dataset before licensing.

Rights are a parallel checklist, not this one. Questions buyers should ask counsel before a training license can fail after the bytes look perfect. What the records are for, if you accept them, is how business data trains AI agents to do real work. Matching of profiles is described on the marketplace page. Matching is not delivery.

Why acceptance is a decision

Teams waste weeks training on a drop that was the wrong version, missing the outcome column, or accompanied by a readme that contradicted the brief. The checklist makes acceptance a recorded decision with named failures. It also gives the sender a concrete discrepancy instead of a mood. A mood cannot be fixed. A missing field can be sent back.

What the receiving team can actually verify

• Formats. The files open in the tools you named. An exotic container you cannot read is a fail, even if the sender prefers it.
• Schemas. Column names, types, and required fields match the brief. Extra columns that contain excluded data are a fail, not a bonus.
• Documentation. A reader who was not in the sales call can tell what a code means and which field is the outcome. Undocumented codes are incomplete.
• Transfer controls. Whatever you agreed: access scope, expiry, who can download. Do not assume a control you did not see. We are not asserting that our marketplace implements a particular control.
• Versioning. The drop is the sample you diligence, or a stated successor. A silent replacement is not the same dataset.
• Integrity. Sizes, counts, or checksums the sender documented match what you received. If they did not document any, you cannot claim you checked integrity.
• Discrepancies. A written list of misses against the brief, sent back before anyone marks the set accepted.
• Acceptance. A named person says the lines passed, or they do not. Silence after download is not acceptance.

What not to assume we provide

Do not read this page as a specification of Big Data/Little Guy's transfer product. Profiles are descriptions. A later handoff, if any, is scoped in the arrangement you actually sign. The buyer overview and the buyer request are how a need is described. They do not deliver archives. How records are limited before a buyer can study them is about limiting a slice a holder approved, which is not the same as your acceptance test.

Handoffs that skip the list

• Training on the drop the day it arrives. You have not compared it with the brief.
• Accepting extra fields that the exclusions forbade because they might be useful later. Extra is a breach of your own brief.
• A checksum you do not recompute. Writing the word integrity in a ticket is not a check.
• No owner for discrepancies. A shared channel where everyone assumes someone else replied.
• Treating acceptance as legal sign-off. The file can match the schema and still be outside the grant. Keep counsel's questions open until they close them.

How we talk about handoff

The operator is Unlockt, LLC, publishing as Big Data/Little Guy. We introduce a stated need to a profile. We do not claim a universal delivery format, a universal checksum practice, or a hosted acceptance workflow. If a transfer is part of a later agreement, the agreement should name the checks. Until then, do not design your pipeline as if those checks already ran.

Holders checking their own data use the free data value check. Buyers describing a need use the request form. If you need the procurement order again, return to the sourcing guide. Acceptance is the last gate, and it is yours.