The Touchline / Club Technology

Why club AI should transcribe less and understand more

Transcription turned speech into text. The next step is turning speech into typed, routed, remembered work — with provenance on every fact.

A performance analyst reviewing match footage on two laptops in a dim video room before dawn

Every club now has transcripts. Meetings are recorded, pressers are captured, training comms are logged. The result, in most buildings, is a folder nobody opens. Transcription turned speech into text; it didn't turn speech into work.

Text is not structure

A transcript of a coach's morning walkthrough contains observations, decisions, instructions, player-specific cues, and thinking-out-loud — undifferentiated. The value isn't the words; it's the type of each sentence:

  • an observation ("their left back steps out early") belongs in the opposition model;
  • a decision ("we start in the diamond") changes the game plan and needs to reach everyone;
  • an instruction ("cut the build-up clips") is a task with an owner and a deadline;
  • a player cue ("tell Kenta to hold the far post") must reach one person, phrased carefully.

Treat all four as "notes" and you've built a diary. Type them, route them, and close the loop on them, and you've built an operating system for the staff.

The graph underneath

The second thing structure buys you is memory that compounds. When "Sota — knee, scan pending" is a typed fact rather than a sentence in a file, the system can answer the questions coaches actually ask: what changed about this player this month? What did we say about this opponent last time? What did I decide, and when?

We model this as a knowledge graph — subject, predicate, object, with a timestamp and a source. Nothing exotic; the exotic part is that it's built passively, from speech, while the coach does their job. No data entry. Provenance on every fact, so "why does it think that?" always has an answer: because you said so, here, on this recording.

The honest constraint

Models mishear names, and a task routed to the wrong person is worse than no task. This is why name resolution — aliases, nicknames, the way a transcript mangles "João" — is the hardest and most important component in the pipeline, and why every automated action in CoCoach carries its confidence and its source. Trust in the building is the real product. The AI is just how you earn it.

aianalysisknowledge-graph