How TacticAI lets football analysts explore a different defensive setup

Football analysis is becoming more interactive. Google announced a TacticAI collaboration involving Palmeiras and the Brazilian Football Confederation in June 2026, extending a programme of work that had previously investigated corner kicks with Liverpool.[1]

Google DeepMind describes a system in which Palmeiras analysts can drag players into different positions and explore defensive setups. It represents the 22 players as a network of connected individuals and models open-play developments up to eight seconds ahead. That is a stated forecasting horizon; the announcement supplies no accuracy figure for those forecasts.[2]

A corner offered a manageable place to begin

The 2024 TacticAI paper used 7,176 Premier League corners from 2020–2021. Player information became a graph: positions, movement, height and weight described the individuals, while connections identified whether players belonged to the same team. The model predicted the first receiver and whether a shot would follow, retrieved similar situations, and suggested defensive adjustments.[3]

This gave the researchers evidence that specialists found the proposed changes useful. It also established a concrete role for generative AI in sport: suggest a different arrangement of people that an analyst can examine.[3]

The picture of the pitch has to be assembled

Earlier work in the same research collaboration addressed a less visible problem. Television footage follows the ball, so players repeatedly leave the picture. DeepMind and Liverpool’s 2022 Graph Imputer study estimated missing player trajectories using observations from before and after the gaps, together with relationships between players.[4]

The researchers evaluated their approach on 105 full-length matches with tracking data, simulating what a camera would hide. They also examined pitch control: which team would be expected to reach different parts of the playing area first. Estimating an unseen defender’s position can change that picture.[4]

That earlier method reconstructed gaps using later observations. It should not be confused with the newer claim of forecasting open play. The connection is a practical one: tactical analysis depends on a representation of the surrounding players, including those the viewer may not have seen.[4]

Palmeiras turned the idea into an interface

In a public account, Matheus de Amorim of Palmeiras describes work spanning data preparation, model development and the interactive simulation interface. He credits technical discussions with TacticAI co-author Petar Veličković. The club’s team adapted the research ideas to its own open-play problem, connecting the data and models to a tool someone could operate.[5]

That account shows a useful division of work. The research contributed a way to represent relationships between players. The club developed an application around a particular analytical question. Preparing the data, adapting the model and building its interface were connected parts of the same project.[5]

Consider a hypothetical defensive review. A coach wonders whether a full-back should move closer to an opponent. That could reduce one receiving option while opening space elsewhere. A simulation gives the discussion an explicit alternative to inspect. The coach can compare the proposed positioning with the original situation, then decide whether the movement fits the team’s players and instructions.

The opportunity is to explore more of these small, consequential choices. The original replay remains the record of what happened; the alternative is material for discussion and practice. An analyst can bring a coach a particular change to consider, with the surrounding movement made visible. That is a useful addition to the craft of preparing a football team.

Report an error or suggest a correction