Why vetting decides data quality
Expert-labelled data is only as good as the expert who labelled it. A title on a profile says what someone has done; it does not show how they reason through a task they have not seen before.
So we test that directly. Each expert is assessed on work that resembles what they will do on an AI project, and the assessment is set for their field.
Stage 1: sourced
We find candidates through four channels: our agentic Hiring OS, outreach agents, a recruiter network and a jobs marketplace. Using several channels means we are not limited to people who already happen to be looking for annotation work.
Outreach agents contact professionals in the fields a project needs. The Hiring OS keeps sourcing, screening and assessment in one pipeline, so a candidate's record follows them from first contact to placement.
Stage 2: vetted
Candidates are interviewed by AI voice agents. The interview asks them to explain how they would approach problems in their field, which shows how they reason, not only what they know.
They then complete real-world work tests in their field and skills assessments. A software engineer might review code for defects; a translator might render a passage with the right register. Credentials and experience are reviewed as part of the process.
When a project needs something more specific, we add it. Vetting can include any assessment type the work requires.
- Interviews by AI voice agents
- Real-world work tests in the expert's field
- Skills assessments
- Project-specific assessments where the work needs them
Stage 3: matched
Passing vetting is not the end. Each expert is placed on the AI project that fits their field and the task, fast.
Matching is specific. A cardiologist and a hospital pharmacist are both medical experts, but they catch different errors. We match on the work, not the category.
What vetting can and cannot tell you
Vetting raises the floor. It filters out people who cannot do the work and shows how the rest reason under realistic conditions.
It does not make anyone infallible. Experts disagree, and careful people still make mistakes, so the data they produce should be reviewed and evaluated like any other data. We say that plainly because buyers should plan for it.
The same process, licensed
The Hiring OS, AI voice agents and outreach agents behind our vetting are technology we built and use every day. Companies that want to run this process on their own hiring can license it.
Licensing gives a company the same pipeline we use to source, interview and assess experts. The method does not change because someone else is running it.