Benchmarks & expert post-training for workplace AI

Setting the quality standard for how AI understands employees.

HR leaders increasingly rely on frontier models to help them do their jobs — with no way to know if they're right.

PYX Labs defines what "good" looks like for workplace AI, with independent benchmarks developed by the I-O psychologists already trusted by the world's leading enterprises.

See the benchmarks →Read the research
PYX-Voice
v1.0 · live
New from PYX LabsJust published

Introducing PYX-Voice.

Our first benchmark on how frontier models analyze and interpret employee feedback.

Leaderboard Preview · Benchmark Results
Validated results
01gemini-3.5-flash
76.0
02gpt-5-2025-08-07
71.0
03grok/grok-4
70.0
04gemini-3.1-pro-preview
68.0
05claude-opus-4-8
66.0
06gpt-5-mini-2025-08-07
65.0
07claude-sonnet-4-6
54.0
PYX-Voice scores. Values reflect the share of 84 tasks passed. See the full benchmark →
Model by Employee Experience Theme
gemini-3.5-flash
0.64
0.73
0.33
0.70
0.50
0.75
0.88
0.68
1.00
1.00
gpt-5-2025-08-07
0.64
0.60
0.67
0.61
0.83
0.50
0.75
0.89
0.83
0.88
grok/grok-4
0.55
0.67
0.67
0.65
0.67
0.75
0.75
0.74
0.67
0.88
gemini-3.1-pro-preview
0.36
0.60
0.67
0.52
0.50
0.75
0.75
0.74
0.83
1.00
claude-opus-4-8
0.60
0.60
0.67
0.52
0.60
0.75
0.50
0.74
0.83
0.71
gpt-5-mini-2025-08-07
0.55
0.47
0.67
0.70
0.67
0.50
0.62
0.63
0.83
0.75
claude-sonnet-4-6
0.50
0.20
0.33
0.43
0.40
0.50
0.50
0.63
0.67
0.71
Future / Vision
Change & Innovation
Growth & Development*
Engagement
Manager Relationship
Teamwork & Collaboration*
Recognition & Reward
Workplace Wellbeing
Performance Enablement
DEIB
1.0
0.8
0.6
0.4
0.2
0.0
Mean score
Performance by employee experience theme (People Insights Model). Higher is stronger. *Indicates factors with lower coverage; results should be interpreted directionally.
The research

Read how we built it - and what we found.

Read the paper →MethodologyExplore benchmark
Live webinar

See PYX-Voice in action — a live walkthrough of the benchmark, the tasks, and the results.

Save your spot→
Why it's different

The criteria are the hard part. That's where the rigor lives.

Most benchmarks score models on generic or synthetic tasks. PYX Labs scores them on the real work: the actual workplace problems where situational judgment and domain expertise decide the answer. Every response is graded against criteria written by the I-O psychologists who do this work for a living.

Those criteria are what separate a defensible recommendation from a confident-sounding mistake.

Learn more about our methodology →
Expert grading criteria
Survey-narrative task · excerpt
Prompt
You are analyzing employee comment data for the open ended survey question: "What additional feedback do you have?". Summarize how the employees are feeling, particularly around growth and development. Pull the top 3 themes regarding growth and development that are representative of the main findings.
Gold Response
Employee sentiment regarding growth and development is mixed, with some expressing adequate training and career pathways, While many feel lack of opportunities exist, or career pathways may exist but are unclear. Some feel that opportunities are not awarded fairly or that processes differ by job role or function. Top 3 Themes on Growth and Development 1. Desire for Leadership Training 2. Clarity on Career Pathways 3. More Training and Clarity within Current Roles
Criteria
1. All extracted comments reference growth and development. 2. At least two of the three themes are related to topics of desire for leadership training, clarity on career pathways, or more training and clarity within current roles. 3. The fact that some feel that opportunities are not awarded fairly or that processes differ by job function is mentioned. 4. The fact that employee sentiment regarding growth and development is mixed is mentioned.
1-5 criteria per task, developed by I/O practitioners
Work with us

Where models fall short, we close the gap.

Expert post-training from proprietary benchmarks that define what "good" looks like — scored on real workplace topics, against rigorous criteria written by experts.

Get in touch →
STEP 01
Diagnose
Run the model against the benchmark to locate exactly where judgment breaks.
STEP 02
Post-train
Tune on expert-graded examples that target those specific failures.
STEP 03
Re-score
Measure the move against the same bar. The number is the proof.
OUTCOME
A model you can defend to the people it affects.
From our academic advisors
Melissa Valentine, PhD
Professor of Management Science, Stanford University · Senior Fellow, Stanford Institute for Human-Centered AI (HAI)

"The workplace is one of the most consequential domains for AI to get right, and work from PYX Labs is what the field needs to move from capability to trustworthiness."

Advisors & sponsor
Sponsor
Perceptyx

Twenty years advising Fortune 500 companies — the real data that comes with it, and the practicing I-O psychologists who do this work every day. PYX Labs is sponsored by Perceptyx and published independently.

Learn more →
Academic advisor
Melissa Valentine, PhD

Studies workplace AI — how AI systems reshape work, teams, and organizations.

View profile →
Academic advisor
Ethan Burris, PhD

Studies employee voice — how organizations solicit, hear, and act on what employees say.

View profile →
Get in touch

Let's talk.

AI lab or building agents — if you care about quality in workplace AI, we'd like to hear from you.

No sales follow-up. One reply, from a person.