A lab setting the standard for quality in workplace AI.
PYX Labs builds the benchmarks that define what "good" looks like for AI doing the work that touches employees — built by the I-O psychologists already trusted by the world's leading enterprises.
Three ways we set the standard.
Benchmarks
Proprietary benchmarks that define what "good" looks like — scored on real workplace topics, against rigorous criteria written by experts.
Expert post-training
The fastest path to move a model toward that standard — tuned on the judgment that separates a good call from a confident wrong one.
Continuous evaluation
The proof the gap stays closed — re-scoring every release against the same expert bar it was tuned to meet.
Backed by the people who built the field.
PYX Labs is a research initiative sponsored by Perceptyx, which contributes a de-identified research archive built over two decades and the practicing I-O psychologists who define what a good answer looks like. The methods and results are published in full to ensure an unbiased benchmark that stands on its own.
PYX-Voice is built by PYX Labs with guidance from academic advisors who study workplace AI and employee voice. Methods and findings are held to academic scrutiny before anything is published.
Studies workplace AI broadly — how AI systems reshape work, teams, and organizations.
View profile →
Specializes in employee voice — how organizations solicit, hear, and act on what employees say.
View profile →Inspect AI is the open-source evaluation framework used by national AI safety institutes. Building on it keeps every run reproducible and auditable — and lets frontier labs inspect exactly how a result was produced.
The rigor is human.
Every task starts as a real practitioner workflow. An expert completes it first — producing the gold-standard answer and the criteria it must meet. Defining what "good" looks like in this domain is the hard part, and it's done by the I-O psychologists who do this work for a living.
Every task and rubric is reviewed and approved by at least two experts before it enters a benchmark. The criteria are what separate a defensible recommendation from a confident-sounding mistake.
Want to be measured against the real work?
Whether you're an AI lab looking to make your model genuinely good at workplace tasks, or building AI agents and need a credible quality bar, we'd like to hear from you. One contact path, one reply, from a person.
Get in touch →