Input
List the source material, required context, permitted references, exclusions, and any condition that should cause a reviewer to stop or escalate.
For employers
AI data services
Define the material, intended use, rubric, versioning, validation evidence, and acceptance questions before discussing AI-data work.
Describe the material and intended use
Start with the material the task will receive and produce: text, code, audio, images, structured records, model outputs, or another defined unit. Explain where the result will be used and which decision it will support. A task for internal analysis may need different evidence and review depth from a task used to change a user-facing system.
Write the task unit
Output
Specify the response shape, required explanation, evidence fields, and how uncertainty or missing information should be recorded.
Rubric
Turn quality into observable criteria. Separate factual correctness, instruction following, relevance, style, safety, and domain judgment when the task needs them.
Examples
Include representative accepted, rejected, and borderline cases. Explain why each example belongs in that category so reviewers can transfer the reasoning.
Version the working definition
Give the task, rubric, and example set explicit version identifiers. Record which version produced each result and why a revision was made. When intended use, source material, or acceptance criteria change, treat that as a new planning decision instead of silently rewriting the old record.
Plan validation evidence
Choose a sample that covers common cases, important edge cases, relevant languages or domains, and any high-consequence failure pattern. Compare results with the current rubric and investigate disagreement. A single aggregate rate can hide systematic failures, so keep the underlying categories visible.
The NIST AI Risk Management Framework is a primary public reference for connecting context, measurement, and risk-management decisions.
Set acceptance questions
- Does the delivered material match the defined input, output, and evidence fields?
- Which rubric version and example set were used?
- What validation sample was reviewed, and which cases remain uncertain?
- What disagreement patterns appeared, and how were they resolved?
- Which findings require task revision before another round?
- Who has authority to accept, revise, or stop the work?
These questions make the first AI-data conversation concrete while leaving scope and delivery decisions to the evidence in the brief.
Bring the brief, not just a title
Share the task context, required judgment, source material, reviewer needs, and the decision your team is preparing to make.