AI Coaching Tools I Actually Used This Season: An Honest Review

Two in the morning, the wind off Table Mountain rattling the windows, and I was sitting in front of three monitors rewatching our team's scrim replays. On one side of the screen, an AI analytics dashboard was drawing heatmaps in real time; on the other, each player's decision-making timing was laid out in graphs. Just a year ago, all of this happened inside my head and in spreadsheets. This season was different. I pushed several AI tools into my actual coaching routine, and some truly changed the way we work while others turned out to be nothing more than pretty graph generators.

What I'm writing here isn't an ad. It's about the things that actually ran on my desk. I've kept only the tools that survived Cape Town's unstable power situation (load shedding struck yet again).

Mobalytics AI Performance Tracker

My expectations were low when I first installed Mobalytics. I'd already tried a few similar services, and most of them stopped at feedback like "your CS is low, place more wards." But the AI coaching feature updated this season was a bit different. It was quite accurate at analysing each player's recent match patterns and catching habitual positioning mistakes.

Our team's support player had a habit of pushing in too deep to secure vision right before teamfights. I told him a hundred times and it never changed, but once I showed him this tool's visualisations, he started improving within a week. It reminded me once again that pattern visualisation is far more effective at convincing players than raw numbers. The downside is that the servers are based in Europe, so data syncing can be slow from South Africa. When you want to review right after an early-morning scrim and loading takes two or three minutes, it breaks the flow.

Attaching GPT-Based Natural Language Queries to Replay Analysis

The most interesting experiment this year was building a pipeline that structures replay data and lets you ask a GPT-based model questions in natural language. For example, if you ask "What do the games where our jungler missed the first dragon in the last 10 matches have in common?", the model sifts through the logs and returns something like "In 7 of those games, wards were prioritised on the top side, and there was insufficient pressure on the bot side."

It's not perfect. Sometimes it mistakes spurious correlations for causation, and if you mess up the data preprocessing, it spits out complete nonsense. Still, it was genuinely useful for figuring out "where should I start digging" before coaching meetings. I used to manually watch three or four replays to find patterns, but now the model narrows down the candidates and I pick only the key ones to examine in depth. Time-wise, it felt like it was cut nearly in half.

Setting it up required writing some Python scripts, and the API costs are not negligible on a monthly basis, roughly R800 to R1,200 per month. But considering the value of the time I'd otherwise spend sitting in front of replays, it more than pays for itself.

Wooting Keyboard Heatmaps and Input Analysis

Strictly speaking, this is closer to hardware-integrated software than an AI tool, but I'm including it because it was surprisingly helpful for coaching. When you visualise the analog input data from a Wooting keyboard, you can see how a player's keystroke patterns change when they're nervous. One particular player pressed keys twice as hard in clutch situations while simultaneously losing input accuracy. Showing them this data and starting with "recognising your tension" had a real mental coaching effect.

Mental coaching used to rely on "feelings" and "conversation," but adding one piece of physical data let both the player and me talk in much more concrete terms. The catch is that equipping the entire team with Wooting keyboards isn't cheap. For now, we take turns using them during practice.

Things That Were Just Noise

I wish everything had been a success, but honestly, there were tools I installed and turned off after just two weeks.

One was a tool that displayed AI-recommended strategies as pop-ups in real time during matches. The idea is brilliant, but in actual scrims, players started waiting for the pop-up instead of making their own calls. What I want to develop as a coach is players' autonomous decision-making, not the ability to follow AI instructions. I turned it off after the second scrim. You know when something sounds perfect on paper but collapses the moment it meets reality? That was this.

Another was a tool that analysed the opposing team's past matches and recommended bans and picks, but its data source included games that were far too old, making its recommendations out of touch with the current meta. I paid about R500 and used it for a month, but I was more accurate just watching the opponent's three most recent games myself.

Looking Toward Next Season

What I've got my eye on now is a prototype that reconstructs replays in 3D within a VR environment, letting players experience "the vision you should have had in that moment." It's not at the commercial stage yet. I'm in talks with the developer on Discord. Sourcing VR equipment in Cape Town isn't easy, so there are practical hurdles, but if this works out, I think it will completely change the way we correct positioning.

I often hear people say AI tools will replace coaches. In my experience, a good AI tool extends the coach's "eyes" but cannot replace the coach's "voice." The shift in a player's expression the moment they watch a replay and realise their own mistake, knowing what to say at that exact moment, that's still a human domain. I use AI precisely to better protect that domain.

Tonight, too, I check the load shedding schedule, verify the UPS battery level, and turn on the monitors. I need to watch three more replays before tomorrow's scrim. The AI has already narrowed down the candidates, so one cup of coffee should do it.