New AI Soccer Analyst Works Step-by-Step With Humans, Not Instead of Them
AI that shows its work could make number-crunching jobs faster and more trustworthy.
Sports analysts have a tough job. They take a question like "why does our team lose leads in the second half?" and turn it into code, charts, and a story a coach can act on. AI language models (software that writes and reasons in plain English) can now do much of that heavy lifting. But there's a problem: you type a prompt, you get a report, and you have no idea how the AI got there. If the answer is wrong, nobody can tell.
That's what this paper tackles. Calvin Yeung and Keisuke Fujii built a tool called AI Soccer Analyst that breaks the work into six visible stages: understanding the data, defining the problem, planning, running the analysis, writing an evidence-backed report, and then refining it together. At any stage, the human can jump in, correct the AI, or send it back. Think of it less like a vending machine that drops a finished answer, and more like a junior colleague who shows you each draft along the way.
The researchers first interviewed five professional analysts to learn what they actually needed. Then they ran a test with 16 participants on 48 real tasks. About 33 of the 48 tasks met the bar for completion — roughly seven in ten — and users generally rated the finished work as high quality, reliable, and verifiable. The logs showed something interesting: human expertise showed up most in the moments where people clarified questions, adjusted the plan, and pushed back. The human wasn't just clicking approve.
The catch is that this is a small, soccer-specific study. Under a third of tasks fell short, and the tool was tested with a handful of people, not a newsroom or a hospital. Still, the design idea travels well. Any job where AI produces an answer that someone must defend — finance, journalism, medicine, law — benefits from software that shows its homework and lets a human steer. Verifiability may matter more than raw speed.
- The tool splits AI analysis into six checkable stages instead of one automatic answer, so you can see and fix each step.
- In a test with 16 people on 48 tasks, about 33 tasks (roughly 7 in 10) were completed successfully.
- Users rated the results reliable and easy to verify — and human experts did most of their work correcting and refining the AI.
Why It Matters
AI you can inspect and correct is safer to rely on at work, where a wrong answer costs money or credibility.