Sports tech's analytics pipeline is fixing lines, not selling you more apps

Sports technology gets judged by its most embarrassing moment. At Wimbledon in 2025 that moment arrived twice in a week. The second one was a ballboy still on court. The Guardian's Sean Ingle reported the fallout on 15 July 2025, and his piece is worth rereading now that the argument has moved on to AI everywhere else.

Two failures, and both were operational rather than algorithmic. An official switched off a line-judging system in one case. In the other, it misread a serve because a person was standing in the wrong place. The system in question was a souped-up version of Hawk-Eye, which Wimbledon has used since 2007. That distinction disappears the moment the story becomes "machines can't do the job".

What the challenge data actually says

Ingle asked IBM how often players got it right when they challenged a line call at Wimbledon in 2024. Of 1,535 challenges across the men's and women's singles, 380 were overturned. That is under 25%. When a player thought the ball was out and made a challenge, they were wrong three times out of four. Separately, researchers had long estimated that line judges got around 8% of close calls wrong.

Those two figures are not a defence of any particular vendor. They are a baseline. The question a sporting body has to answer is not whether a system is perfect, because nothing is, but whether it beats what came before and whether it is improving. Hawk-Eye is more accurate now than in 2007. That is the shape of the argument, and it is a duller shape than the outrage.

There is a second reason governing bodies keep buying: protecting officials from abuse. At the last Rugby World Cup, referee Wayne Barnes described receiving threats of sexual violence against his wife and threats of violence against his children. Whatever you think of automated calls, that context is real, and it is not going away.

The human baseline is worse than we admit

Then there is the research on how people actually decide. One study asked 40 qualified football referees to judge 47 incidents from a Liverpool versus Leicester match. Half watched with crowd noise, the control group in silence. The group hearing the crowd awarded significantly fewer fouls against Liverpool, at 15.5%. A separate Norwegian study found successful teams were more likely to get favourable penalty decisions. Psychologists call it conformity influence. Machines do not have it.

"No system is 100% perfect, but they are demonstrably more accurate than relying purely on human decision-making," said Matt Drew, who founded the integrity department at StatsPerform.

Drew's point is the pragmatic one. Sports use the technology because it gets more decisions right and shields officials and players from abuse, and the best deployments, tennis and cricket among them, keep the fan experience intact. The NFL is replacing its chain gang with Hawk-Eye from September. The Premier League has been watching how Fifa ran VAR at the World Cup and Club World Cup, with fewer delays and the same replays shown to fans.

Where the analytics money is moving

None of this is confined to line calls. Agentic analysis is the current phrase for letting software chase a question through several datasets without a human writing every query. Starburst's September 2026 write-up on two architectures for agentic data analysis sets out the trade-off plainly: hand an agent raw extracts and you pay to ship terabytes it must then process itself, or give it direct database access and let it write and refine its own queries while the engine does the heavy lifting. The second option is the one that scales, though Starburst notes agents can be more demanding than human users and can flood a system with speculative queries.

Replication is the other half of the problem, and it is being measured. A CData white paper published through Wiley's science and engineering hub in September 2026 reports a 76% speed-up in large-table replication on the same infrastructure, using parallel partitioned reads that split a source table into row ranges read simultaneously across CPU threads, plus write-path optimisation that cuts per-file and per-column overhead. Wide tables with hundreds of columns gain most, because the per-column work repeats on every file operation. The numbers come from a vendor with something to sell, and the methodology is in the paper rather than the headline, so treat the 76% as a vendor benchmark rather than an independent result.

Open source is filling in the same layer from the other end. Adaca Analytics, published on GitHub in September 2026, runs self-hosted Google Analytics 4 and BigQuery dashboards on Cloudflare Workers, with daily rollups landing in a D1 database the operator owns and realtime views staying on Google. Plainoldanalytics, also on GitHub, embeds web analytics directly in Go applications and flushes traffic to disk roughly once a second. Neither is a sports product. Both are the kind of plumbing that a club's data team now reaches for before it buys anything.

That is the quieter story underneath the Wimbledon headlines. The visible failures get a fortnight of columns. The accuracy figures, the replication benchmarks and the self-hosted dashboards get a changelog. One of those is compounding, and it is not the outrage.