The Future of Work Runs on Energy, Not Time

Five years ago, a good chunk ofyour job was execution: writing the first draft, building the first version,pulling the numbers together. AI does most of that now, often before you'vefinished your coffee. What's left on your plate is the part nobody's automatedyet: deciding whether the draft is good enough, whether the numbers tell thereal story, whether the plan is worth acting on. It’s the same job descriptionand outputs, with a completely different workflow.

McKinsey's most recent analysis of US workputs a number on it: 57% of current work hours could theoretically be automatedwith existing technology, and the shift underway is less about jobsdisappearing than about people moving from performing tasks to directing them,framing the problem, interpreting the output, deciding what happens next.

There are two separate storiestangled up in that shift.  The first is:what happens to your own brain if you let AI carry too much of the thinking?The second is: how are you set up for the job that's left when you use AI well?Because it demands good judgement and that’s a far more expensive fuel thanexecution ever was. The goal was never to make AI more human. It's to freepeople to focus on what humans do best, and that only works if both we’re ableto find the right balance.  

The Skill You Lose

In June 2025, MIT Media Labresearchers Nataliya Kosmyna and Eugene Hauptmann ran the study that manysceptics of AI are pointing too. Fifty-four participants wrote essays acrossthree sessions in three groups: one using ChatGPT, one using a search engine,one using nothing but their own brain, all wired up to EEG. The pattern wasblunt. Brain connectivity scaled down in a straight line with how much externalhelp a group had. Unassisted writers showed the strongest, most widelydistributed neural networks. The ChatGPT group showed the weakest.

In the recall test, minutesafter finishing, most of the ChatGPT group could not accurately quote a singlesentence of the essay they had just written, and reported the lowest sense ofownership over work that didn't quite feel like theirs.

Then came the twist. In a fourthsession, months later, the groups swapped. Participants who'd used ChatGPTthroughout wrote unassisted, and their neural connectivity came back weakerthan the group who'd been writing unassisted from the start. Participants who'dbeen unassisted and were then given ChatGPT showed improved recall instead.Order mattered: build the capability first, and AI sat on top of it fine. Skipstraight to AI, and the capability wasn't there to fall back on. Worth sayingplainly, since the study itself says as much: this is onepreprint, not yet peer-reviewed, and 54 participants is not a huge sample. Butit's the first research to watch this happen inside the brain in somethingclose to real time, and every result points the same way.

The Job That's Left Anyway Costs More to Run

Here's the separate problem, andit exists even for someone whose critical thinking hasn't dropped at all.Execution work is repeatable. You can do it tired, do it on autopilot, do it at40% capacity and still get an acceptable result. Judgment work isn't like that.Deciding whether a model's output is right, spotting the plausible-soundingerror, holding a strategic view when three tools suggest something different,all of that draws on sustained attention and pattern recognition that runs downfast and doesn't come back with a coffee. It’s the energy that it took to restructure this blog, using Claude, 3times.  

A Microsoft Research study found the twoproblems can feed each other: the more confident knowledge workers were in AI'soutput, the lower their own critical thinking scores. Confidence in the tooland capacity in the person moved in opposite directions, at exactly the momentthe person's capacity was what mattered most. Dr Talia Varley's answer, in Fast Company, is to treat cognition itself asoccupational health rather than a personal responsibility to manage alone:protect deep thinking time the way you'd protect physical safety, and measurewhether people still have the capacity to disagree with a machine, not justwhether they're keeping up with it.

Two Different Problems, Same Deadline

Worth being precise about whichproblem is which, because the fix differs. The MIT findings describe atrophy:what happens to capability you stop exercising. The energy problem describescost: what it takes to run the capability you still have. You could keep yourcritical thinking fully intact and still burn out from judgment work stackedback to back. Or you could pace your energy perfectly and still lose theunderlying skill by handing every hard decision to a model instead ofpractising it yourself. Most organisations manage only one of these, usuallythe energy one, because it shows up on a productivity dashboard. The atrophyrisk shows up months later, with no symptom to point to, in decisions that usedto be obvious and no longer are.

Energy Management Is the New Time Management

For a hundred years, managing work meant managing time:blocking hours, hitting deadlines, filling a calendar efficiently. That madesense when jobs were built from execution, which scales roughly with hours putin. Judgment doesn't scale that way. Your capacity for it moves with yourbiology: the sleep you got, where you are in a stress or recovery cycle,whether your prefrontal cortex has anything left to give by 4pm.

That's the real shift underneaththe AI headlines. The scarce resource at work is no longer time. It's theenergy required to do the part of the job that's still yours, and the practicerequired to keep that judgment muscle from going quiet. Every knowledge workeris now managing an energy budget and a capability budget at once, and mostcalendars, task lists, and performance reviews still only track the first one.

Redesigning the Job, Not Just the Tools

None of this gets solved by oneperson managing their own energy harder while the job stays the same. Theorganisations getting this right are redesigning roles around where judgmentnow sits, protecting both the energy it costs and the practice it needs. Callit brain-positive leadership: protecting deep thinking time like physicalsafety, building in unassisted reps of judgment work so the skill doesn'tatrophy unnoticed, and watching for the same strain signals you'd watch on a burnout dashboard: decision fatigue, reduceddebate, quiet over-reliance. Burnout has always measured brain strength undersustained load, not a personal failing. AI just changed what's doing theloading, and what's at risk of going quiet from disuse.

How to Use This with Phase

Phase was built for the energyhalf of this equation. Connect a wearable and Phase reads your sleep score,readiness, or cycle phase, then ranks your task list by bio fit rather thandeadline, so the judgment call only you can make lands in the window when yourbrain has the most to give. AI-assisted first drafts and routine admin getpushed to the stretches where your energy is lower, because that work canabsorb it without the quality dropping.

Sync your calendar and the samelogic runs a level up: a week stacked with high-stakes review sits against youractual energy forecast, so you can see a collision, a big decision on theworst-slept night of your month, before it happens rather than after.Protecting cognitive capacity means two things at once: thinking at the momentyour biology can support it, and still doing enough of the thinking yourselfthat the muscle is there when you need it.

AI was never going to make workeasier in a straight line. It's changing what work is, both by costing youcapability if you lean on it too hard, and by leaving behind a job that's moreenergy-intensive than the one it replaced. Treat both risks as real, not justthe one on a dashboard, and you're still doing good judgment by Fridayafternoon, and still capable of doing it without a model's help.

Phase already tracks the energyside. Join the waitlist and start planning your weekaround your actual energy, not just the hours you've got left in the day.

‍