Smarter than the Analyst, Dumber than the Coach

There is a lot of noise about AI in sport right now.

Every conference has a panel on it. Every performance director has an opinion on it. Vendors are lining up to tell you their platform will revolutionise the way you coach. 

It can be exhausting — and if you are anything like me, your instinct when something gets hyped this hard is to take a step back and ask what is actually true.

So let me tell you where I am coming from before I tell you anything else.

I have always had a bit of an obsession with finding better answers. Not because I think I am smarter than anyone else in the room, but because I genuinely believe that the players and teams I work with deserve the best available information — and I have never been comfortable with the idea that we might be leaving useful insight on the table simply because we did not have the tools or the processes to surface it. 

That restlessness has driven a lot of decisions in my coaching career. It is what pushed me toward video analysis when most coaches were still relying purely on memory and instinct. It is what drew me to GPS and physical load data when those tools first became available. And over the past two years, it is what has pulled me deep into the world of artificial intelligence.

I will be honest with you: it started as a fascination. It has become something closer to a discipline — one I am still learning, still testing, and in no way claiming to have mastered. But I have done enough work with these tools in a live coaching environment to have formed some strong views about where they genuinely help, where they can mislead you badly, and what it takes to use them responsibly. That is what this piece is about.

I am not writing this as a technology evangelist. I am writing it as a working coach who has spent a significant amount of time getting his hands dirty with these systems and believes the rugby community deserves a more honest conversation about them than it is currently getting.

The title is deliberately a little self-deprecating. AI, used correctly, is faster than any analyst on your staff at processing data, finding patterns across large datasets, and holding dozens of variables in mind at once. But it is, for now, considerably worse than any experienced coach at reading a room, understanding a player’s emotional state, managing a team dynamic, or making the instinctive decisions that define great coaching. The relationship between AI and the coaching craft is not a competition. It is, when done right, a collaboration — and the terms of that collaboration matter.

What we actually mean when we say “AI”

When most people in rugby circles talk about AI, they are imagining something from a science fiction film — a machine that watches your match footage, thinks like a tactician, and hands you a game plan on Monday morning. That is not what this is. Not yet, and perhaps not ever in that form.

The AI that is genuinely useful to coaches today is a powerful engine for processing, querying, and interpreting structured data. It is a system you can instruct, feed information, and that will find patterns, surface anomalies, and generate outputs at a speed and scale that no human team can match. Think of it less as an oracle and more as the most diligent, tireless assistant you have ever worked with — one who has read everything, forgets nothing, and will never complain about working through the night before a Test match.

How you think about the tool shapes how you use it. Coaches who approach AI expecting magic will be disappointed and, worse, may be misled. Coaches who approach it as a powerful analytical engine — one that still requires human judgment at every step — will find it genuinely transformative.

“AI is information on steroids. That can be incredibly useful, or it can be deeply misleading. The difference lies entirely in how you use it.”

The process: from raw data to coaching intelligence

The work I have been involved in runs on a principle that sounds simple but demands rigour: every meaningful action in a rugby match can be coded, stored, and interrogated. The game generates an extraordinary volume of events — carries, tackles, line breaks, defensive resets, set-piece sequences, kick contestations, ruck outcomes, exit plays — and for years, coaching teams have been capturing this information in spreadsheets. The problem has never been collecting data. It has been doing something genuinely useful with it, at speed, without introducing bias or error into the process.

The workflow moves through several distinct stages, and each must be completed with integrity before moving to the next.

THE PIPELINE

01  Match coding.  A trained analyst codes each match event by event, assigning values to action type, player, team, field zone, phase, and outcome. This is human work. The quality of everything that follows depends entirely on the accuracy of this step.

02  Database construction.  Coded data is imported into a structured database, with raw data preserved untouched. Separate cleaned and standardised tables are built, and a data dictionary maps every column, every value, every transformation. Nothing is assumed. Everything is documented.

03  Validation and quality assurance.  Before any analysis begins, the data is checked for errors. Row counts are reconciled. Missing values are flagged. Unexpected categories are investigated. Duplicates are found and resolved. This step is not optional.

04  AI-assisted analysis.  Only once the data has passed validation does the AI engage. Queries are built to surface trends, calculate metrics, compare performance against benchmarks, and identify patterns that would take a human team days or weeks to find manually.

05  Human interpretation.  The outputs are reviewed by the coaching team, cross-referenced against lived match experience, and translated into practical training and selection decisions. The AI presents findings. Coaches make decisions.

What this process allows is something genuinely new: the ability to hold months of match data in a single queryable environment and ask questions that would previously have been impossible in any practical timeframe. How has a specific area of our game trended across eight matches? What does the data say about our performance under pressure in the final twenty minutes? Which patterns are linked to winning and losing? The answers are in there — and now we can reach them.

The most important person in the room: your analyst

Before we go any further into what AI can do, I want to say something that does not get said often enough: none of it — not one piece of insight, not one player report, not one trend analysis — is worth a thing if the underlying data is wrong.

The foundation of any AI-assisted performance system is a human being sitting in a coding suite, watching match footage, and assigning a value to every single action. That person — your analyst — is not a support role. They are not a technical afterthought. They are the most important person in the room. Everything flows from their accuracy. Everything.

Think about what coding actually involves. An analyst must make a judgment call on every event in the match — the action type, the player involved, the field zone, the phase, the outcome. A single systemic error in how a category is being coded — say, an inconsistency in how a defensive breakdown contest is classified — will propagate silently through the entire database and corrupt every metric that touches it. The AI will process it without complaint, present the results with complete confidence, and the coaching team will make decisions based on numbers that are wrong. Not approximately wrong. Precisely, convincingly, dangerously wrong.

This is why the coding process must be governed by clear, written definitions that every analyst in your organisation applies identically and consistently. It is why inter-rater reliability checks — where two analysts code the same passage of play independently and compare results — are not optional extras but a core quality assurance discipline. And it is why the raw coded data must be preserved untouched at every stage, so that any anomaly identified later can be traced back to its source.

“Artificial intelligence does not fix bad data. It amplifies it. Feed it inaccurate input and it will return inaccurate output — confidently, fluently, and at scale.”

If you need evidence that the world’s best rugby minds understand this, consider what happened recently. Rassie Erasmus — the architect of back-to-back World Cup victories and arguably the most strategically astute coach in the game today — appointed Joe Lewis as the Springboks’ new performance analyst. Lewis is Welsh, holds a master’s degree in performance analysis from Cardiff Metropolitan University, and spent the last nine years as England’s head and senior analyst, most recently on Steve Borthwick’s staff through the 2026 Six Nations.

Think about what Rassie is actually acquiring here. Not just an analyst. A living intelligence database — nine years of England’s internal systems, preparation philosophies, analytical frameworks, and institutional knowledge, walking through the door of the Springbok set-up. And he is doing it not in place of his existing analysts, but on top of them. While other programmes debate whether to hire an extra coach, South Africa is investing in additional intelligence. That distinction tells you everything about how Rassie thinks and why the Springboks continue to pull away from the rest of the world.

Analytical capacity is not an administrative function. It is a competitive asset. When you lose it, you feel it.

Your data, your identity

There is a temptation, when starting this kind of work, to reach for what is readily available. Several large sports analytics companies offer comprehensive data products — match statistics, player benchmarks, league-wide performance comparisons — and the appeal is obvious. The data is clean, structured, and arrives without the overhead of building and maintaining your own coding operation.

That data has legitimate value for contextual analysis — understanding where your team sits relative to global standards, tracking competitor tendencies, benchmarking players. I am not dismissing it.

But I want to be clear about what generic external data cannot give you, because I think this is one of the most under-discussed risks in the growing rush to integrate AI into rugby performance environments.

External data reflects how rugby is played on average. It captures what is common. It describes the mean. If you feed that data — unfiltered, undifferentiated, without the context of your own team’s identity — into an AI system and ask it to define what good performance looks like for your players, you are not building a performance culture. You are outsourcing one. You are allowing a generic description of average rugby to become the benchmark by which your people are assessed and developed.

Every team that achieves sustained success does so because it has a clearly defined way of playing — principles, priorities, and non-negotiables specific to that group and that coaching philosophy. The standards that matter are not the league average for defensive tackles made. They are the standards your coaching team has defined, your players have bought into, and your system has been built around. Those standards live inside your organisation. They do not arrive in a data feed from a third party.

The player at the centre

Of all the applications of AI in this work, the one I feel most strongly about is its potential to serve individual player development. This is where the technology stops being abstract and becomes something that directly affects a young player’s career, confidence, and growth. Done well, it is one of the most powerful tools a coach has ever had. Done badly, it can do real harm.

Consider the traditional approach to player feedback. A coach watches footage, takes notes, draws on memory and intuition, and sits down with a player to share observations. There is enormous value in that conversation — the human connection, the trust, the nuance of tone and body language. But it is limited by time, by memory, and by the cognitive impossibility of tracking every relevant action for every player across every match in a season.

AI-assisted reporting allows us to hold all of that information simultaneously and present it in a form that is accurate, consistent, and tailored to the individual. A player’s kicking performance, for example — contestation outcomes, territorial gains from different field zones, success rates under pressure versus in space — is the kind of multi-variable picture that a spreadsheet simply cannot present cleanly. An AI-assisted individual report can.

The report does not tell the player whether he is a good kicker. That is a coaching judgment. What it does is give both the coach and the player an honest, data-grounded picture of where performance is strong, where it is inconsistent, and what the numbers suggest should be prioritised in training. It converts abstract impressions into specific, actionable observations. And it does so with the same rigour and consistency for every player in the squad — no unconscious bias, no selective memory, no favouritism.

“The best player development tool is one that gives a player the truth about themselves — specific, fair, and impossible to argue with.”

The design principles are worth stating explicitly. Every metric must have a definition and a denominator — not raw counts, but rates and ratios that account for volume and context. Small samples are flagged, not hidden. Outputs are presented as observations, not verdicts. And the report always ends with a training implication — a bridge from data to action that makes the information useful rather than merely interesting.

The emotional intelligence of delivery remains entirely human. AI produces the report. The coach delivers it. That distinction is not a limitation of the technology. It is a feature.

The dangers: a word of caution

I want to spend real time on this, because I think it is where the rugby community — and sport more broadly — is most at risk. And I would say this especially to younger coaches who are beginning to explore these tools: please read this section carefully.

The most seductive and dangerous thing about modern AI systems is that they sound authoritative. They speak in full sentences, with appropriate hedging, with an air of having considered the evidence carefully. They do not sound like they are guessing. And that is exactly the problem — because sometimes they are.

When you ask a general-purpose AI model a question about rugby — about a team’s tendencies, about a player’s reputation, about what style of play is most effective — the model draws on everything it has ever processed. That includes not just academic papers, coaching manuals, and broadcast analysis, but also newspaper opinion columns, fan forum debates, social media threads, and the passionate but sometimes wildly inaccurate assessments of supporters who watched the same game you did and reached entirely different conclusions.

Rugby generates enormous volumes of emotionally charged opinion. Supporters feel deeply about their teams and express those feelings loudly across the internet. AI systems that learn from this content cannot reliably distinguish between a rigorous analytical observation and a supporter’s frustration after a heavy defeat. Both inform the model’s understanding. When the model synthesises that understanding into a confident-sounding response, it may be giving you the statistical average of ten thousand opinions — some expert, some not — presented as though it were fact.

“The internet gave us access to all the world’s information. AI gives us access to all the world’s information, at speed, with an editorial voice. That is extraordinary. It is also, in the wrong hands, extraordinarily dangerous.”

The scenarios that concern me most: using a general AI model to inform game-planning decisions without verifying the underlying data. Using AI-generated assessments of opposing players as the basis for selection conversations. Or — most dangerous of all — presenting AI-generated player feedback to athletes without having rigorously verified every number and observation against your own validated source data.

If a player is told that his defensive metrics suggest he is below standard in a particular area, and that assessment is based on poorly validated data or on opinion rather than coded match events, the damage to that player’s confidence and the coach’s credibility can be profound. Player trust, once broken through inaccurate feedback, is very difficult to rebuild.

WHEN AI IS USED CORRECTLY	WHEN AI GOES WRONG
• Validated, internally coded source data • Every metric defined and denominated • Quality assurance before any output • Human interpretation at every stage • Findings presented as observations, not verdicts • Small samples explicitly flagged • Coach delivers — AI informs	• Generic questions to general- purpose models • Opinion-sourced data treated as fact • No validation layer before analysis • Outputs trusted without corroboration • Findings used directly in player feedback • Source data quality assumed, not verified • The tool replaces judgment rather than informs it

A framework for getting it right

Given everything I have described — the extraordinary potential and the genuine risks — what does responsible AI use look like for a coaching team in practice? Based on our experience, I would offer the following principles as a starting framework. These are not abstract ideals. They are the rules we work to every day.

Final thoughts

I began this piece as a convert, and I remain one. The work we have been doing with AI-assisted match analysis and player reporting has genuinely changed how I coach, how I prepare, and how I think about the information available to a modern coaching team. There are things I know now, from the data, that I could not have known before — at least not quickly enough to act on them effectively.

But conversion does not require abandoning critical thinking. The most dangerous person in any coaching environment is the one who has found a new tool and trusts it unconditionally. Rugby is a game of relentless complexity, played by human beings of extraordinary variability, in conditions that no model can fully capture. The data tells us important things. It does not tell us everything.

My honest advice to any coach considering how to engage with AI: start with your own data, build rigorous processes, maintain your scepticism, keep the player at the centre of everything — and never, under any circumstances, present an AI-generated finding to a player without being confident enough in it to put your name to it.

Because ultimately, that is what coaching is. Not the analysis. Not the data. Not the tool. A person standing in front of another person, with their reputation and their relationship on the line, trying to help them get better. AI, used well, gives you more to work with in that moment. The moment itself remains entirely yours.

That is why the coach, for now at least, is still smarter than the machine.

Over to you I have described what AI-assisted analysis can do in principle. In the next piece, I want to show you what it looks like in practice. Send me a genuine rugby question you would like answered using real match data and an AI-assisted analytical process. I will choose one request, take it through the full process — coding, database construction, validation, AI-assisted analysis, and coaching interpretation — and publish the results as a worked example. No black boxes. No hand-waving. The full pipeline, from raw question to coaching insight. Send your question: rugbyiq.com@gmail.com

GG

 

Thank you to Cheridan Inglis and her team at INtouch Strategy for creating our marketing communications.