AI Coaching in Esports: iTero, GIANTX and the Grey Zone Nobody Has Defined

Opening: twelve minutes with no rulebook

In a best-of-five series, the most dangerous stretch of time is not on the map. It sits between game two and game three, lasting somewhere between eight and twelve minutes, when the head coach and the analyst have to decide what they will change before the next game starts. Inside that window there is a laptop. On that laptop there may be a spreadsheet, a locally running machine-learning model, or a paid analytics platform. No rulebook of any major league states clearly what is allowed to run on that laptop, and nobody has defined the boundary between a tool that supports thinking and an opinion produced by a machine.

The interview with Jack Williams about iTero, GIANTX and the future of AI coaching sits exactly on that intersection. If the arithmetic drawn from the source text itself is right, the piece was written around 2026: it recalls Natus Vincere lifting the Aegis of Champions at Gamescom fourteen years earlier, and that title dates to 2026, at the first The International. That detail is personal nostalgia, but it gives me a hard anchor point for everything else. And once I had the anchor, I started counting how much genuinely verifiable information the article contains.

Context: an interview that contains very little match data

I approach every document the same way before writing: I count how much information belongs to the subject and how much belongs to the author. This time the result forced me to change how I read. Of the thirteen information points I could extract from the source, ten describe the article's own author, referred to by the nickname Ollie, along with his ambition to recreate the Natus Vincere run at Gamescom. Only three points concern the interview's actual subject, and two of those are confirmed only at heading level, with no body text behind them.

Put plainly: the article contains no patch data, no bracket structure, no player names, no win rates, no pick-ban figures. Anyone writing that a given patch made a given team stronger on the basis of this document is inventing. I once made exactly that mistake. In 2026 I built an entire pre-match argument for South Korea against Iran in World Cup qualifying on two metrics alone, expected goals and progressive passes. The match finished 0-0, the team only secured its place on the final matchday, and I sat back down with thirty-eight qualifying matches from five confederations to redo the analysis from scratch. That mistake taught me that data never lies, only the reading of it does. Afterwards I built a rule I do not break: every claim needs at least two independent layers of data and a note on its margin of error.

The two section headings disclosed in the source are working exclusively with Giant X and the likelihood of being copied, plus the topic of AI-assisted cheating. A commercial frame on one side, an integrity frame on the other. Between them sits a third frame the article never opens: competitive fairness within the league. These three frames do not exclude one another. They stack, and the order in which they stack determines whether an exclusivity deal is ordinary business or a governance problem.

I do not trust intuition, I trust numbers that speak once they have been asked the right question. Here, though, the right question lies precisely where there are no numbers at all. That is why I am writing this as a map of gaps rather than a set of conclusions.

Core: which window the tool runs in, and who gets to plug it in

First, an AI coaching tool has to be split by time window, because each window carries a different level of legality. The first window is pre-match: preparation, opponent analysis, draft planning. The second is between games within a series: tactical adjustment, priority shifts, processing newly emerged information. The third is post-match: review and data building for the next round. The fourth is in-game, in real time.

The fourth window closed long ago in every major title. Real-time assistance is explicitly prohibited, so there is nothing left to debate there. The genuine grey zone sits in the second window, the break between games in a best-of-three or best-of-five. There, a model can read the data from the two games just played, cross-reference a historical corpus, and produce a recommendation before the coach finishes a sentence. No rule forbids it, and no rule explicitly permits it either. My confidence in this claim is medium, because the source does not confirm which window iTero's product operates in.

The second point is league structure. GIANTX, to the best of my industry background knowledge, is an EMEA-rooted organisation competing in the LEC, formed through the merger of Excel Esports and Giants Gaming. Confidence here is medium and requires verification, including the possibility of a different entity sharing the same rendering of the name. If that information is correct, the governing framework for the iTero-GIANTX arrangement is Riot Games' third-party software and competitive integrity rules.

The notable part is this: in a closed franchised league, structural advantages are not competed away season by season. The LEC has no relegation. Its members are permanent members. A team holding exclusive access to an analytics tool keeps that advantage across multiple seasons rather than seeing the market level it. In an open system with promotion and relegation, weak teams drop out of the competition entirely, and a tooling advantage is naturally bounded by the short life of a competitive slot. In a closed league, tooling advantages compound. This is why I hold that exclusivity carries far greater structural consequences in a franchised league than in an open circuit.

The third point, and in my view the most overlooked variable: patch cadence determines the value of an AI model. Valve ships large systemic updates infrequently, creating long stretches of stability between changes. Within that stability a model trained on historical data retains its validity, and that validity compounds over time. Riot Games runs a different rhythm, shipping patches continuously through the season. That rhythm shortens the half-life of every learned pattern.

The consequence is concrete. In a title with a stable patch cadence, AI is a knowledge asset. In a title with a fast patch cadence, AI is a tempo asset. A knowledge asset deepens with use, and exclusivity means something. A tempo asset holds value only for a short window before opponents reach the same conclusion on their own, and exclusivity means almost nothing. A single product marketed identically across both kinds of titles is, to my mind, a signal to ask questions about the measurement methodology behind it. My confidence in this chain of reasoning is medium, since it rests on the known release rhythms of the two publishers rather than on any performance data from iTero.

The fourth point is the measurement vacuum. The source discloses no sample size, no evaluation methodology, no win-rate differential before and after adopting the tool. That is not necessarily the vendor's fault. The whole industry lacks a shared standard for talking about coaching-tool effectiveness. I have seen the same script play out in football with expected goals. It took years for competing data providers to even sit down together, and to this day they publish different values for the same shot, because their underlying definitions differ. A metric without a shared definition cannot be used to adjudicate fairness.

The fifth point is the copying question. Between the numbers of any transfer is a story nobody puts in the report. The same applies here: the commercial terms of the iTero-GIANTX arrangement are undisclosed, so any judgement about contract value is speculation. But I believe the fear of copying is aimed at the wrong asset. Model weights are not an economic moat. Weights can be bought, retrained, or replaced by an open-source model that is good enough. The harder thing to copy is the process of questioning the model: which variables to select, how to normalise them, how to weight context, and who holds veto power over machine output. That process lives in people, not in files.

Based on my experience following matches in the LEC and at Major-level Dota 2 events, I notice a difference in decision rhythm. In leagues with dense patch cycles, coaching staffs tend to talk about confirming a feeling, meaning they have already reached a conclusion and are looking for data to reinforce it. In leagues with sparse patch cycles, they talk about searching for what they do not yet know. Those two psychological states require two different kinds of tooling. They also require two different kinds of contract, a distinction the esports technology transfer market does not currently make.

One detail about betting markets is worth raising here. The betting market is not wrong; it merely reflects a truth you have not yet noticed. Bookmakers react extremely fast to player transfer news, yet price tooling news almost not at all. If an exclusive tooling deal genuinely produces a few percentage points of win-rate difference, that is an information gap sitting in public view for weeks. This claim carries low confidence, because no performance data has been published to verify it, and I offer it as a hypothesis to track rather than a conclusion.

The counter-intuitive angle: the biggest risk is not cheating

Most public debate circles around whether AI is helping people cheat. I think that is the wrong frame, and the wrong frame causes the industry to miss two larger risks.

First, banning AI tools is close to unenforceable. A model running on a personal laptop leaves no trace on a league server. There is no way to distinguish a coach reading a manual stat sheet from a coach reading model output, if both retype their conclusions by hand. What can be detected is not the tool but an invalid input during the between-games window. That means any future regulation will have to target the data flow rather than the software. And that is a far harder technical problem than writing a prohibition clause.

Second, the risk of homogenisation. If every team in a closed league adopts the same tool, or tools trained on the same public corpus, their outputs will converge. Drafts will look alike. Lane priorities will look alike. Stylistic diversity, the thing that gives esports its spectator value, gets compressed into a single local optimum that everyone finds at once. Cheating breaks the fairness of one match. Homogenisation breaks the appeal of an entire season. Over the long run, the second is harder to repair.

Here I have to remind myself of a familiar trap. The cancelled 2026 Seoul derby is a stress test for every prediction algorithm, and it taught me that anomalous events expose a model's limits faster than any validation set. A disruptive event inside the coaching-tool ecosystem, say a publisher abruptly changing its data access policy, would collapse every performance assessment built before it. I have seen no sign of that in the source, but it is a scenario that belongs on the table in advance.

Closing: three signals for the next cycle

I will be tracking three things over the coming season. One, whether a league operator forces tooling agreements to open equal access to all members, or lets exclusivity stand. Two, whether any vendor publishes an evaluation methodology with sample sizes and confidence intervals, turning tool performance from a claim into verifiable data. Three, whether the between-games window is written explicitly into regulation instead of being left blank as it is now.

Esports does not need luck; it needs people who read the meta faster than the servers do. This time the meta is not in a patch. It is in a rule file nobody has written yet, and in twelve minutes that nobody has taken responsibility for defining.