Acts of Evolution

Our take

Bubble Watch: AI Has an All Star Team and a Rec League

Citadel's AI turns six to eight weeks of PhD level research into hours. Most offices are still learning the playbook, and the gap is coaching, not talent.

Acts of Evolution ยท

Around the league takes

The quick box score

Why it matters to you: The AI spending boom assumes every office plays like the best team in the league. The box score says most of the league has barely practised.

Ken Griffin has seen the highlight reel of the future.

Griffin spoke at the Goldman Sachs Apex Symposium in June, as reported by The Motley Fool and Hedge Fund Alpha. The Citadel founder described an agentic AI system inside his firm that can read an academic finance paper, rebuild its method, check the published results and test them out of sample. Time on the clock: roughly two to three hours per paper.

The same play used to take teams of master's and PhD level researchers six to eight weeks.

The Motley Fool reports that earlier this year Griffin said seeing advanced research work handled by AI left him "fairly depressed," because of what it could mean for highly educated knowledge workers.

That is a real result. It is also not the game most people are playing.

The analytics department vs the rec league

Citadel is the team with the giant analytics department. The people who built that research agent were already technical. The people using it are pros who know the sport and the equipment.

Most companies are a Tuesday night rec league. Same ball, same court, very different results.

That gap is not mainly about the models. It is about coaching.

The London School of Economics and Protiviti found AI users save an average of 7.5 hours a week. The asterisk: 68% of employees had no AI training in the past 12 months, so a lot of that gain is left on the bench.

The AI tax is a turnover problem

Then there are the turnovers.

BCG's 2026 AI at Work survey found close to half or more of respondents spend more time reviewing and correcting AI output or managing and directing AI.

Think of a point guard who puts up big numbers but coughs up eight turnovers. The headline stat looks great. The game film tells a different story. Once you count the human work needed to make the output usable, the productivity dividend shrinks.

Practice reps are the binding constraint

Several independent studies point at the same weak spot on the roster.

Wharton and GBK Collective (2026 enterprise AI study). 80% of leaders report positive returns from AI. The biggest split between bosses and staff was training. 78% of leaders said employees get adequate AI training. Only 56% of entry and mid level employees agreed.

LSE and Protiviti. Trained employees were far more likely to use the tools, 93% versus 57% without training. They also saved roughly twice as much time, about 11 hours a week versus 5. Training mattered more than generation. A trained Gen X employee got more productivity from AI than an untrained Gen Z employee. Reps matter more than birth year. Every coach already knew that.

Boston Consulting Group (2026 AI at Work survey). 42% of regular frontline AI users said they save eight hours a week, a full workday. Yet 66% got limited or no guidance on what to do with that time, and more than half were not putting it into more strategic work. The league handed out new sneakers and never changed the playbook.

Why the tracker reads this way

The AI spending boom rests on one big assumption. Powerful models will spread quickly into broad productivity across the economy.

So far the spread is real but narrow. Technical, well supported teams post huge numbers. The average office worker, short on time and light on coaching, mostly gets small assists.

That does not mean the models are weak. It means the models are no longer the main limit for a lot of knowledge work. The limits are workflow redesign, clear guidance on how to use saved time, steady training and whether non technical staff can find the hours to learn the tools. Better equipment does not fix a team that never practises.

At Citadel, the gains widened what an already elite team could chase. "There's no reduction to headcount at Citadel on the back of this breakthrough," Griffin said, as reported by Hedge Fund Alpha. That works for a franchise that drafts for technical fluency. It is not how most organisations run.

The final whistle

For the next phase of the AI cycle, the stat that matters is not another benchmark jump. It is whether companies can close the gap between the small technical squad posting big numbers and the much bigger group of office workers who are not.

That means treating education and process change as the main event, not the halftime show.

Until then, a lot of the big talk about fast, economy wide productivity will keep running ahead of the scoreboard. The tools keep getting better. Most of the players still lack the time, the coaching and the support to use them well.

That gap is the real Bubble Watch.

Get the Pulse in your inbox every week: subscribe on Substack.

Sources