People talk about AI in coaching as if it's one thing. Either a magic box that writes your sessions, or the thing coming to replace you, depending on who's talking. It's neither, and treating it as a single thing is why so much of the conversation goes nowhere.
AI in coaching isn't one capability. It's five, stacked on a ladder, and each rung does something the one below it can't. Knowing which rung a tool sits on tells you exactly what it can do for you and where it'll let you down.
Here's the part most people get wrong before we even start. Climbing the ladder does not mean the AI takes over more of your job. The human stays in charge at every single level. What changes from rung to rung isn't who decides. It's how much the tool prepares before you do.
Level 1: the generator
The bottom rung, and the one nearly everyone has already met. You ask, it produces. "Give me a U16 session on fast breaks," and a U16 session appears. It's a one-off. It doesn't know your team, doesn't remember the last thing it made you, and will cheerfully give you a different answer if you ask the same question twice.
This is where almost every AI tool a coach has tried actually lives, ChatGPT included. It's genuinely useful for getting unstuck. On its own, though, it's a generator, not a coaching system, and it's the rung where you have to check everything, because the tool is working from your sentence and nothing else.
Level 2: the copilot
A step up. Instead of just handing you a finished thing, it helps you work through a task. Compare two versions of a session. Tidy up a plan you've drafted. Lay out three options and let you pick.
You're still the one driving. The copilot rides along and makes the task quicker. More useful than a plain generator, but still waiting for you to point it at something. Nothing happens until you start it.
Level 3: the contextual assistant
This is the first real jump, and you feel it straight away. The tool now knows your actual situation. Who's available this week. What you've got coming up. So when three players drop out, it doesn't write you a generic session, it adjusts the one you had for the squad you'll actually have on the night.
That's the move from answering a prompt to working from your team. It's also the rung most tools promise and quietly miss, because knowing your real context is much harder than writing fluent text. Here you approve the change instead of building it from scratch.
Level 4: the persistent assistant
Now it remembers. Not just this week, but the shape of your season. It connects this session to the last one, notices that your transition defence has fallen apart in the closing minutes three games running, and brings it to you as a pattern rather than a one-off.
Here's the claim worth stopping on. This rung matters more than a smarter model does. A brilliant AI with no memory is still a level-1 generator in a nicer coat. A plainer one that remembers your season will out-help it every week of the year. When we put a generic frontier model up against our own system in testing, the thing the generic one couldn't reproduce wasn't intelligence. It was memory: the player and match history that lets a tool hand you a next step tied to your actual season instead of generic advice. Continuity beats raw cleverness. At this level you validate the read, because interpreting a pattern is still a coach's call.
Level 5: the bounded agent
The top rung, and the one with the scary name. An agent can prepare or carry out approved actions, draft the updated calendar and queue the message to players, not just talk about it.
The word doing the real work there is bounded. This isn't autonomous AI running your club while you sleep. It acts only inside the permissions you set, and approval stays explicit. It does the last stretch of admin once you've said yes, and not a step before. Take the science fiction off the word "agent" and what's left is narrow and supervised on purpose.
The thing that doesn't change
Read back down the five levels and watch the human each time. You check, you direct, you approve, you validate, you set the permissions. The ladder describes how much the tool prepares. It says nothing about how much it decides, because that answer is the same at every rung: the coach decides. Decision authority is flat all the way up. Only the preparation gets deeper.
That's the bit the "AI is coming for coaches" story misses completely. A higher rung is a tool doing more of the legwork, not more of the thinking.
Where AI in coaching sits today
Hold the ladder up to what's on the market and most of it clusters at the bottom. Generators and copilots, levels 1 and 2, dressed up in different interfaces. The value for a club lives higher, at 3 and 4, where the tool knows your team and remembers your season.
So the leap that matters isn't a cleverer generator. You can have the smartest level-1 tool on earth and it'll still forget you the second you close the tab. The jump worth making is upward, to something with context and memory.
Where this leaves you
Put your own tools on the ladder. If everything you use is a level-1 generator, that isn't a small gap, it's the whole opportunity. The piece we built first, the planning, is the context-and-memory part of the ladder rather than another generator. Pick your age group and see it at squadx.app/plan-builder.
If you want the companion to this, we've written separately about the four jobs a good assistant should actually do across these levels. And the full report sets out where AI in handball genuinely is today, with the sources behind it.
