It took more than 30,000 repetitions before the brain moved a single task out of the bottleneck. Georgetown neuroscientists trained participants to sort morphed car images into two categories, tracked them with fMRI and EEG before and after, and found that the work only migrated out of the prefrontal cortex — the region that handles one thing at a time — and into the temporal cortex after five to ten weeks of daily practice (Georgetown University Medical Center, 2026). Before that point, the task consumed executive capacity. After it, the same task ran almost free.
That finding has an uncomfortable corollary for anyone running AI tool adoption in a mid-market operation. A new tool is, by definition, not automatic. It sits in the prefrontal bottleneck and spends exactly the cognitive capacity you deployed it to free — and it stays there for weeks. The productivity dip your team reports in month one is not resistance, and it is not a sign the tool was the wrong pick. It is the predictable cost of a skill that has not yet been rerouted.
What Georgetown Actually Measured
The study, published in the Journal of Cognitive Neuroscience and led by first author Patrick Cox with senior author Maximilian Riesenhuber, is unusual because it is longitudinal. Most learning research captures either the first hours of skill acquisition or the finished expert. This one measured the same brains before and after an extended training block, using a phone app so participants could accumulate more than 30,000 trials over five to ten weeks (Georgetown University Medical Center, 2026).
The result: early in learning, categorization lit up the prefrontal cortex. After extended practice, the same categorization was happening in the temporal cortex, and the category signal bypassed the prefrontal cortex entirely, connecting straight to output regions (Neuroscience News, 2026). As Cox put it, extensive training "essentially put a category-selective area in the temporal lobe that was not there before."
Riesenhuber's illustration is the one every operator already knows from their own life: driving. When you first learn, it takes your whole attention. Years later you can hold a conversation and think through a problem while doing it. The brain did not get faster at deliberating about driving. It moved driving out of the place where deliberation happens.
Two facts from that architecture are the ones that matter operationally. First, the prefrontal cortex is a genuine bottleneck — it can typically only handle one demanding task at a time. Second, the escape from that bottleneck is not a matter of understanding, motivation, or training-day quality. It is a matter of repetition volume, measured in weeks.
The Un-Automatic Tool Spends the Capacity It Promised
Now map that onto how AI tools actually enter a mid-market workflow. You license a tool because a workflow is eating too many hours. The business case says it frees capacity. Then you hand it to a team that has never used it.
For that team, every interaction with the new tool is a prefrontal-cortex task. Deciding what to delegate to it, phrasing the request, judging whether the output is right, catching where it is subtly wrong, deciding whether to accept or redo — all of that runs through the same single-lane executive resource the tool was supposed to relieve. The old workflow was, for an experienced operator, substantially automatic. It had already made the migration Georgetown measured. The new one has not.
So the first weeks are not a fair trade of "old effort out, new efficiency in." They are a trade of automatic work for deliberate work — a straight increase in executive load, even when the tool is objectively good and the output is objectively faster. That is the mechanism behind the dip. Not resistance. Not change fatigue as a personality trait. Architecture.
Behavioral data is consistent with this. ActivTrak's 2026 analysis of 443 million hours of activity across 163,638 employees found that after AI adoption, average uninterrupted focus sessions fell roughly 9% — from 14 minutes 23 seconds to about 13 minutes — while email activity rose 104% and business-tool usage rose 94% (ActivTrak, 2026). Work got faster and denser. The attentional container it runs in got smaller. That is what the transition period looks like from the outside.
Why Stacking Three Tools Triples the Load, Not the Output
Here is where the neuroscience turns into a budgeting error with a price tag.
If one un-automatic tool occupies the bottleneck, three un-automatic tools do not share it gracefully. The prefrontal cortex does not parallelize; Georgetown's finding is precisely that true parallel processing becomes possible only after one of the tasks has been rerouted out of it. Deploy three tools in the same quarter and you have not tripled your leverage. You have tripled the demand on the one resource that cannot be split, while extending the time before any of the three reaches automaticity — because attention spread across three novices' curves accumulates repetitions on each one more slowly.
The mid-market is already living this. Freshworks' 2026 Global Cost of Complexity study, surveying 12,021 IT decision-makers with more than 9,000 in mid-market organizations, found companies running an average of 4.2 AI tools, with 9% running seven or more — and losing an average of 25% of their AI budget to complexity before seeing a single return (Freshworks, 2026). In the same study, 86% of IT leaders said managing AI complexity increased their team's workload.
That 25% is usually explained as an integration and configuration problem. Part of it is. But the Georgetown result names a second cost that no integration budget captures: the human bottleneck tax of running several un-automatic tools at once. Configuration overhead is visible on an invoice. Prefrontal saturation shows up only as a team that is somehow busier after you gave them time-saving software.
The compounding is what makes this expensive rather than merely slow. A tool that reaches automaticity stops drawing on executive capacity and starts returning it — that returned capacity is what funds the next adoption. A portfolio of tools that all stay just short of automatic never produces that dividend, so every subsequent deployment is financed out of the same depleted account. Four-point-two tools is not a sprawl problem in the procurement sense. It is a compounding problem running in reverse.
Sequence AI Tool Adoption for Automaticity, Not for Coverage
Most AI rollout plans are sequenced for coverage — get the tools in front of as many teams as possible, as fast as procurement allows, because the pilot budget expires. The Georgetown mechanism argues for sequencing by automaticity instead, and it changes three concrete decisions.
Fund weeks, not launches
If the reroute takes five to ten weeks of dense repetition, then the meaningful unit of an AI deployment is not the launch date — it is the repetition window that follows it. A rollout that trains a team on Tuesday and moves the change-management attention elsewhere on Wednesday has funded the most expensive part of the curve and abandoned the part where the payoff lives. Budget the follow-through: protected usage time, a named owner, and enough volume of real work through the tool that repetitions accumulate daily rather than weekly.
One workflow at a time, deliberately
Run a single tool through a single workflow to automaticity before opening the next front. This feels slower and is faster. Two sequential deployments that each reach automaticity beat three simultaneous ones that all stay in the bottleneck — and the first one, once rerouted, actually frees the executive capacity the second one will need. Georgetown's practical promise is that the capacity does return. It just returns on the far side of repetition, not on the far side of a training deck.
Stop reading the week-one dip as a verdict
The most expensive misreading in AI tool adoption is treating the early dip as evidence about the tool or the team. Leaders who see it and conclude "our people are resisting" respond with more mandates. Leaders who conclude "this tool doesn't work" churn to a new vendor — which resets the clock and puts the team back at trial one. Both responses are the same error: mistaking a transition cost for a verdict. The relevant question in week three is not "are we faster yet" but "are repetitions accumulating." Measure usage depth and frequency early; measure hours saved after the window closes.
There is a targeting layer worth naming. The reroute is repetition-driven, but the tolerance for the deliberate-effort phase is not evenly distributed across a team. Some people will push through a five-week trough without flinching; others will quietly revert to the old workflow in week two and take the repetition curve with them. Knowing which of your people are which — before you pick the pilot group — is the difference between a deployment that reaches automaticity and one that stalls at 40% adoption and gets written off as a tooling failure.
The One Decision This Quarter
Count the AI tools your team has been asked to learn in the last ninety days. If the number is more than one per workflow, you are not measuring an adoption problem — you are measuring a queue at a single-lane resource that cannot be widened by asking people to try harder.
So pick one. Choose the highest-value workflow, commit a single tool to it, protect the repetition for five to ten weeks, and hold every other AI tool adoption behind that line until the first one runs without deliberation. That single sequencing decision costs nothing, requires no new vendor, and is the only intervention in this article that acts directly on the bottleneck rather than around it. The capacity you were promised is real, and Georgetown just showed the physical mechanism that delivers it. It arrives on the other side of repetition — and only for the teams that were allowed to finish.
