Your AI Rollout Automated a Task. That's Why Nothing Changed.

In March 2026, ActivTrak published what happened when AI landed inside 1,111 organizations. They tracked 163,638 employees. Email time went up 104%. Chat and messaging went up 145%. Focused, uninterrupted work fell 9%.

Read that again. Every communication metric more than doubled. The one metric that actually correlates with good work went backwards.

That is the real state of most AI rollouts right now. Not failure. Something more awkward than failure: a lot of visible activity sitting on top of work that nobody redesigned.

Busier Is Not The Same As Better

The pattern is easy to reproduce. You buy seats. You run a training session on prompting. You tell people to use it where it helps. People do exactly that, and they use it on the tasks directly in front of them, because that is the only surface they can see.

So the drafting gets faster and the summarizing gets faster, and then all that extra output has to go somewhere. It goes into more email and more messages, which lands on somebody else's desk, who now has more to process. The tool worked. The workday got worse.

David Brooks gave this a name in The Atlantic in June, borrowing ActivTrak's data to describe what he called AI brain fry. His larger worry is that AI sorts people into those who think harder and those who quietly stop thinking. He is pointed about the cause though, and this is the part worth keeping: he does not think it is inevitable. He thinks institutions decide it.

The Thinking Doesn't Vanish. It Moves.

There is a study that should be far more famous than the ones that get quoted. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of using generative AI at work. Peer reviewed, published at CHI 2025.

Two findings matter.

The first: confidence in the AI predicted less critical thinking. Confidence in yourself predicted more. Same tool, opposite outcomes, and the variable is the person using it.

The second is the one I keep coming back to. Generative AI does not remove the thinking. It relocates it. The effort moves off information-gathering and onto verification. Off problem-solving and onto integrating whatever the model handed back. Off doing the task and onto stewarding it.

If the skill moved, then handing someone a licence teaches them nothing. They are now doing a job they were never trained for, using judgement nobody described to them, on output that is wrong just often enough to matter. Of course focus time drops.

What Uber Did Instead

In July, Uber's CTO Praveen Neppalli Naga described how his company approached this, and the mechanism is more interesting than the headline numbers.

Uber took roughly 30 of its most AI-fluent engineers and embedded them directly with domain experts in finance, legal, HR and marketing. Each pairing runs a two-week sprint on a fixed shape: days one and two shadowing the expert's actual workflow, day three prioritizing, days four and five building, days six through nine validating, day ten shipping.

It scaled to 16 pods across 16 functions in two months.

The results Uber reports: capital allocation across 150 cities went from 15 hours to 30 minutes. Financial pacing reports went from two days to 10 minutes. Marketing web QA went from two weeks to roughly 50 minutes. Around 9,000 manual support workflows moved to self-service.

Naga's own summary of the lesson is the sentence I would put on the wall: "The biggest wins rarely come from automating one task. They come from rethinking an entire workflow."

Look at what the pod structure actually solves. The engineer knows what the technology can do but has no idea how capital allocation really works, including the informal parts nobody documented. The finance expert knows the work in their bones but cannot see which parts are now cheap to automate. Neither one produces the result alone. That gap is the entire problem, and Uber's answer was to close it physically, by putting the two people on the same workflow for ten days.

Citi Ran The Same Play, Much Wider

Citi went at it from the other direction. Rather than a small embedded team, they named over 2,000 colleagues as AI champions and accelerators, explicitly to drive adoption in day-to-day work and feed a loop back to the people building the tools. By 2026 that cohort was around 4,000, with their internal toolset in front of roughly 180,000 employees and more than 80% using it regularly.

The structural detail worth stealing: it is two tiers, not one. A small senior champions cadre of about 25 to 30 people, and a much larger accelerator pool. Different jobs. The small group sets direction and standards. The large group does the spreading.

Neither company bought their way here. Both built capability into named people and gave those people a job.

Four Moves That Travel To A 40-Person Company

You do not need Uber's headcount for any of this. The mechanism scales down further than people assume.

Pair fluency with domain depth, in the same room. Your most AI-capable person and the person who actually owns the process, on one workflow, for a fixed window. Do not send either in alone. The whole result lives in the pairing.
Shadow before you build. Uber spends 20% of a ten-day sprint just watching the work happen. That feels indulgent right up until you notice that the steps people describe in a meeting and the steps they actually perform are different, and the gap between them is where the automation breaks.
Pick a workflow, not a task. If the win you are chasing is "this takes 40 minutes and now it takes 10", you are optimizing a step inside a process that no longer needs to be shaped that way. Ask what the whole sequence would look like if you designed it today.
Teach verification, not prompting. Prompting is a weekend skill. Judging whether output is right, knowing which claims to check, knowing when to throw the whole thing out, that is the actual job now, and almost nobody is being taught it. Make your champions model it out loud.

The Honest Read

Every Uber number above is self-reported by their CTO, published days ago, audited by nobody. Citi's figures are first-party too. Treat both as directional.

The cognition research needs the same discipline. The MIT Media Lab study everyone cites about AI and your brain is a preprint with 54 participants, and the authors themselves reject the brain rot framing it gets used for. It is a signal, not a verdict.

What survives the scrutiny is the mechanism, not the metrics. Two companies with nothing in common decided that capability is something you build into named people rather than something you purchase, and both of them made the AI-fluent person sit beside the person who owns the work. That part is cheap to copy at any size.

The tools are not the constraint anymore, and they have not been for a while. The constraint is that nobody has been given the job of redesigning how the work flows. Give somebody that job.