The most common question about ScoutRoom is some version of "so the AI writes the scout?" No. The model does the grind. I do the read. That split isn't a limitation I'm working around; it's the design, and the research says it's the right one.
The argument from both failure modes
Humans and models fail at scouting in opposite directions, and both failure modes are documented.
Human scouts see few games, anchor on stars and reputations, and overweight the performance that happened in front of them. Every coach knows the player who got signed off one great night against us. Eyes are precious and biased, and there are never enough of them.
Models fail differently. Train a model on historical selections and it learns whoever the old system favoured; the machine-learning literature is blunt that a model predicting "who gets picked" has learned the bias, not the talent. And every serious study hits the same ceiling: fit, character, durability, motivation, the response to a hostile road crowd. No stat line carries them. The peer-reviewed governance work on AI talent tools (Morgulev and Azar, 2026, Big Data and Cognitive Computing 10:146) lands on a principle worth keeping verbatim in mind: machine scores belong in screening and prioritisation, and the final call needs a documented human review. Víctor Orta, then Sevilla FC's sporting director, put the practitioner version on record at the World Football Summit in 2023: never sign a player on data alone, and never sign one without it either.
So the field's answer is settled, and it's "both": the model for breadth, the human for judgement. What varies between products is which half gets respected.
What each half actually does here
The model side of ScoutRoom does what a model is good at and a person at 11pm is terrible at. It compiles every box score, every game, both teams, no fatigue. It normalises, ranks, flags: this team's turnover pressure is top of the conference, this shooter's volume is a starting-guard number, this rotation shortened three weeks ago. Breadth, consistency, arithmetic that never gets sloppy in the fourth quarter of a long night. It also carries the receipts: every number stamped with its sample and source, every tendency carrying a confidence grade.
Then the part a model can't do. Which of those twelve flags actually changes how you guard them on Saturday? What does this coach do after a timeout when the first option is taken away? Is the drop in a big's minutes load management or a lost rotation spot? What will this specific bench actually execute, with these players, on two training nights a week? That's coaching judgement, built from years of being the person the plan fails on when it's wrong. The teardown that ships is the model's breadth passed through that filter, with a name on it.
Why this is a promise, not a caveat
"Human in the loop" gets used as a disclaimer, a hedge against the machine's mistakes. I mean it as the opposite: it's the part you're paying for. Anyone can generate a stats dump now; the marginal cost of unread numbers has fallen to zero, which means an unread number is worth roughly what it costs. What a club buys from ScoutRoom is that someone with skin in the coaching game read everything, threw most of it away, and signed the four calls that survived.
The model widens the net. The scout still signs the player, and the scout still signs the scout. If a teardown is wrong, there's a person who was wrong, on purpose, with their reasoning on the page. Model progress only makes the grind cheaper. The judgement stays the whole job.