● WORKFLOWS WE RUN
Data labeling workflows
See how people can prepare and check the information AI learns from. These are representative ways we can support a project. The exact work, tools, decision rights, and quality checks are agreed with each customer.
01 / AI DATA
Multilingual audio transcription
Create accurate written records from conversations for speech and language projects.
WHAT WE START WITH
Audio recordings, language requirements, and transcription instructions.
WHERE PEOPLE CHECK
People check speaker turns, unclear words, numbers, and formatting according to the project guidelines.
01 / Set instructions
Agree on languages, formatting, and how to handle unclear speech.
02 / Transcribe
People listen to audio and create or check transcripts.
03 / Check
Review samples, resolve disagreements, and correct errors.
WHAT YOU RECEIVE
Reviewed transcripts in the agreed format, plus agreed notes about issues requiring a decision.
02/ AI DATA
Label text and documents
Add useful categories or tags to text so it can be used in AI projects.
WHAT WE START WITH
Text or documents, categories, and labeled examples supplied by the customer.
WHERE PEOPLE CHECK
Reviewers apply the agreed categories and flag unclear or overlapping cases.
01 / Define
Agree on the categories and examples.
02 / Label
Reviewers read and tag each item using the instructions.
03 / Check
Check consistency and clarify confusing cases.
WHAT YOU RECEIVE
Tagged records and a log of questions or issues, as agreed for the project.
03 / AI DATA
Evaluate AI answers
Review AI-generated answers using criteria provided by your team.
WHAT WE START WITH
AI answers, evaluation criteria, and examples of good and poor responses.
WHERE PEOPLE CHECK
Reviewers assess the answers against the rubric and flag cases where the rules are unclear.
01 / Agree
Define what a good answer looks like.
02 / Assess
People review answers against the criteria and record feedback.
03 / Check
Review disagreements and update guidance as needed.
WHAT YOU RECEIVE
Recorded assessments and agreed feedback, ready for the customer to review.
04 / AI DATA
Compare AI responses
Help teams understand which of two answers better meets their criteria.
Build the foundation
Prepare and review the data that helps teams train and evaluate AI systems.
Run the workflow
Manage the review, exceptions, and follow-through that turn an AI output into completed work.
01 / Set criteria
Define what reviewers should prioritize.
02 / Compare
Reviewers compare responses and explain their choices where required.
03 / Calibrate
Discuss difficult examples and check for consistency.
Improve together
Use quality checks and operational feedback to refine guidelines and identify recurring issues.
FRONTIER DATA / 01
Capture a workflow trace
Create an approved step-by-step example of how a person completes a business task. A trace helps an AI team study the process, not just the final result.
1. Define the task
Select an approved scenario, starting materials, and success criteria—for example, reconciling a sample invoice.
2. Capture the work
Record permitted steps, information viewed, decisions, tool results, and relevant artifacts.
3. Review and package
Check completeness and remove or de-identify restricted information, then deliver the agreed trace and metadata.
FRONTIER DATA / 02
Collect decision and preference data
Gather structured human judgment about competing options, with enough context to understand what made one choice better.
1. Examine the run
Review the task, available action history, tool responses, and final outcome.
2. Annotate the issue
Mark the step where a problem occurred and categorize it, such as a missed check or incorrect tool input.
3. Suggest a correction
Record a better next action, explain it against the rubric, and escalate unclear cases.
FRONTIER DATA / 03
Collect decision and preference data
Gather structured human judgment about competing options, with enough context to understand what made one choice better.
1. Set the criteria
Define the decision, relevant context, and standards for a good result.
2. Compare or decide
Reviewers choose among responses or actions and provide a concise rationale where requested.
3. Check agreement
Audit samples, resolve differences, and prepare judgments for training or evaluation.
FRONTIER DATA / 04
Build tasks to test an AI agent
Turn realistic business scenarios and exceptions into examples that help teams evaluate whether an agent can complete work reliably.
1. Create the scenario
Prepare an approved task and supporting documents, including incomplete or unusual cases.
2. Define success
Describe expected outcomes and a rubric for accuracy, appropriate escalation, and task completion.
3. Review results
Assess the agent’s output and available steps; document errors, ambiguity, and cases needing expert review.
Examples are illustrative and can be tailored to each project. Access, collection rights, consent, privacy safeguards, and delivery formats are agreed before work begins.
HOW WE WORK TOGETHER
Built around your process.
We start by understanding the work, your requirements, and what a completed task should look like. Then we agree on instructions, review checkpoints, responsibilities, and reporting before the team begins.
-
Start with a defined batch or workflow and review the results together.
-
Agree on how questions, unclear cases, and sensitive information will be handled
-
Use feedback and recurring issues to improve the instructions and the process.
Scope, tools, service levels, measures, and deliverables are tailored to each engagement.