Which AI should run your admin?
First test run is queued
6 assistants · 12 tasks · 42 automatic checks. Results publish here automatically. Every prompt is below so you can see exactly what we’ll ask.
Read every answer
Pick a task to compare the exact responses, check results and grader notes.
Write a short quote email from Tomás at Greenline Landscaping to a homeowner named Priya. Line items: new sod, 400 sq ft at $2.10/sq ft; old lawn removal, flat $450; labour, 12 hours at $75/hour. Show each line total, the subtotal, 5% GST, and the final total. Keep it under 150 words and end with how to accept the quote.
Responses appear here after the first run.
How the TAIT score works
Identical prompts
Every assistant gets the same 12 prompts word for word, through its official API, with default settings and no system prompt.
Automatic checks (30%)
42 pass/fail tests for instruction-following: word limits, required formats, valid JSON, banned words, facts kept. No opinions involved.
Blind rubric grading (70%)
Each answer is graded on its own, without the product name, against 4 fixed criteria per task. When two grader models are available, scores are averaged to reduce bias.
Re-run on schedule
The suite re-runs automatically so scores track model updates. Every run is versioned and the full raw data is kept.
Limits, stated plainly: AI graders can share blind spots with the models they grade, and a 12-task suite can’t cover every use. That’s why every output is published next to its score. Some links on this site are affiliate links; they never affect scores, which are computed by the code in our public test suite.