Every Team Got the Same AI. That Is Why the Data Platform Decided Who Won

One detail from this tournament’s technology programme should worry anyone whose AI strategy is “we bought a good model”: FIFA made a generative AI analysis platform available to all 48 teams. Same tool, same capability, delivered to the eventual winners and to the sides eliminated in the group stage.

That is the clearest natural experiment in enterprise AI I have seen. Identical model access, wildly different outcomes. Whatever separated the teams that got value from it, it was not the model.

The commoditisation nobody priced in#

The strategy decks from two years ago assumed model access was the moat. Get the better model, get the better outcome. That assumption has quietly collapsed, and not just in football.

July alone saw GPT-5.6 land in three variants, Grok 4.5 ship, Meta’s Muse Spark 1.1 arrive, DeepSeek V4 go stable on the 24th, and Kimi K3’s open weights promised for the 27th. Capability that was frontier in January is a commodity API call by July, at a fraction of the price. Anyone can rent it. Your competitor rented it last week.

When the model is the same for everyone, the differentiator moves to the two things that are not portable: the data you can feed it and the workflow it plugs into.

What a team actually needs before the model helps#

Consider what “AI-assisted match analysis” requires to be useful on the day before a knockout fixture.

The opponent’s last six matches, tracking data included. A Premier League match reportedly generates around 3.2 million data points; a full tournament’s worth across 48 squads is a serious volume. That data arrives from multiple providers in incompatible formats, with different player identifiers, different event taxonomies, and different coordinate systems — some vendors put the origin at the corner flag, some at the centre spot, and getting that wrong silently mirrors your entire analysis.

It needs to be joined to your own squad’s readiness data, which lives in a completely different system owned by the medical staff.

It needs to be current. Analysis of a team’s shape that does not account for the centre-back who got suspended two days ago is worse than no analysis, because it is confidently wrong.

And it needs to land in front of a coach in a form they will act on, within a window measured in hours, in the middle of a tournament where the staff are travelling between cities.

The model does none of that. Every one of those is ingestion, identity resolution, freshness, and delivery. Which is to say: it is a data platform problem with a model on top, and the model is the last and easiest 10%.

The four things that separate teams that get value#

We see the same four determinants across clients, in industries that have nothing to do with sport.

Entity resolution that actually works. If the same player, patient, student, or customer has three identifiers across three source systems and no reliable mapping, every downstream analysis is subtly wrong and nobody can tell you by how much. This is unglamorous, it is where the effort goes, and skipping it is the most common reason AI implementation projects produce outputs that domain experts distrust on sight.

Freshness measured in the units the decision needs. Not “we refresh nightly” as a technical fact, but “the decision is made at 16:00 and the data is 20 hours old at that point.” Write down the decision time and work backwards. Most pipelines are built to a cadence nobody derived from a decision.

A place to put the answer. The single most predictable failure mode is a good analysis that lands in a channel the decision-maker does not read. A coach who gets insight in a PDF emailed at 23:00 is a coach who does not get insight. This is the same reason we push analytics into the operational system rather than a separate BI portal — a Hospital Management System insight belongs in the bed manager’s screen, not in a weekly deck, and a School ERP early-warning flag belongs in the form tutor’s morning register.

Someone who can tell when it is wrong. The teams that got value from a generative analysis tool had analysts who knew the sport well enough to recognise a plausible-sounding hallucination about an opponent’s press. That expertise is not replaced by the tool; it is what makes the tool safe to use. Deploy a model into an organisation with no one qualified to challenge its output and you have automated the production of confident errors.

The human layer under the numbers#

There is a detail here worth sitting with. A lot of the structured event data feeding these platforms is not generated by computer vision at all — it is produced by human annotators. Reporting from this tournament found data workers in the Philippines, Brazil, Cambodia, India, Egypt and Ukraine tagging matches by hand, capturing up to 3,000 actions per game across three to four hours of work, for figures around 60 euros per match in one Rio case.

That matters for two reasons. First, it is where the ground truth comes from, so annotation quality is a hard ceiling on model quality — a point worth remembering the next time someone proposes fixing accuracy by upgrading the model. Second, it means the “AI” stack has a labour supply chain with real people in it, and pretending otherwise is both inaccurate and, increasingly, a procurement question your clients will ask you about.

What I would tell a CTO who just bought a model#

Three things, in order.

1. Audit your identifiers before you audit your model. Pick your three most important entities and count how many distinct identifiers each has across your systems. If the answer is more than one and there is no mapping table with an owner, stop and fix that first. Nothing downstream is trustworthy until you do.

2. Write down the decision and its clock. Which decision does this improve, who makes it, at what time, and what is the current data age at that moment? If nobody can answer, you have bought a capability without a use.

3. Assume the model advantage is temporary and the data advantage compounds. Every model you can access, your competitor can access within a quarter, cheaper. The joined, resolved, fresh dataset of your own operations is the only asset in the stack that does not depreciate to zero on someone else’s release schedule.

Forty-eight teams, one tool. The tool was never the variable.

---

If your AI programme is a model in search of a pipeline, you are paying frontier prices for commodity output. We build the plumbing that makes the model worth having. Start with the data.