The work after the work gets faster
Working with AI is exhilarating and exhausting. I’ve automated away most of my menial work and am performing at levels I once couldn’t imagine. At the same time, I’m working harder than ever and coming home more mentally drained than before. Even how I think about the nature of collaboration is changing.The mundane work gave me breathing room between stretches of high-octane strategic deep work. Where I used to feel lucky to make progress on one hard thread, I can now make significant progress on several concurrently. That’s exhilarating. It also means thinking harder and holding more of those demanding streams in my head at once. The exhaustion comes from both.
First drafts are ready in minutes. Thorough analyses are as simple as asking questions out loud. I can explore with a breadth, depth, and precision that previously took a small army of expensive experts. Articulating excellence is enough to create excellence in the vast majority of knowledge work.My favorite quote is Arthur C. Clarke’s Third Law: “Any sufficiently advanced technology is indistinguishable from magic.” (opens in a new tab) Let’s take a beat to recognize that we’ve entered an era where people can create legit software and build successful businesses by speaking conversationally with an artificial intelligence. Sounds like casting magic spells to me.
And yet … the machine got faster, but I didn’t. Nobody handed me a new job description. The old one still stands, and something new has been stacked on top of it that I’m only starting to be able to name.
The shape of the work changed
Making gets faster. The work changes shape.
Effort concentrated in making
Most of the effort goes into producing the work.AI accelerates making
More of my effort goes into direction, judgment, and review.Coval (opens in a new tab)’s CEO, Brooke Hopkins (opens in a new tab), often talks about the inversion of engineering work, and I think the same principle applies to many other areas of knowledge work.In her February 24 post on systems that keep improving (opens in a new tab), Brooke argues that getting to a demo has become easier while the harder work is building the infrastructure to evaluate, learn, and improve safely after launch. The diagram here is my interpretation of the shift in human effort across knowledge work. The engineers I work with have had more time and a wider range of experience with AI tooling than many revenue leaders. I look to them for patterns we might carry into other knowledge work. The words and mental models may differ, but questions about planning, testing, review, and ownership are relevant across revenue, finance, operations, design, people, and legal. Each function still needs to work out what those patterns mean in its own context. A useful way to picture the shift in workflows I’ve observed is an inverted U becoming a U. Before, the effort clustered around making: a little planning, a big hump of making, and a little refining on the tail end (what engineers would call testing). With AI, more of my effort goes into the two ends: planning up front and refining afterward, with much less time spent making in between.
In some of my workflows, making happens hundreds of times faster. Review and delivery haven’t accelerated at the same rate. As I produce more, the work of deciding what’s useful and getting it into use grows too.
A brief is cheap to generate and expensive to act on. I can generate a detailed one in seconds and still hand someone an expensive assignment. If they have to read the source documents, check the summary, and reconcile the differences, the workflow eats more attention than it saved. Beautiful formatting is no longer a signal of quality or usefulness. Even if every brief I produce is a succinct half-page primer on something genuinely useful, I can generate them far faster than anyone can act on what I’m finding.
Here’s one of mine. A demo request comes in, and an agent immediately reads the prospect’s LinkedIn profile and their company page, works out potential needs and use cases, drafts a pre-meeting email, drafts a list of questions, and assembles a customized pitch deck. Amazing, right?! I built those skillzI use “skillz” with a z for agent skills: reusable instructions and resources that help an agent perform a particular task. “Skills” with an s are human capabilities, developed through learning and practice. The spelling helps me keep the two distinct when talking about how they grow together. and I stopped using them, because it was all too much. I had to skim the website and LinkedIn myself to gauge whether the synthesis was in the right ballpark. Then I had to read each deliverable carefully to see if it felt and sounded right, editing as I went when something was off from how we’d say it. I’d arrive at the meeting genuinely prepared, the conversation would take a left turn after slide two, and most of the prep would evaporate.
I’m now redesigning that workflow around a short brief with pre-baked action recommendations. The goal is to reduce the cognitive load of deciding what to do next. That decision is quickly becoming a bottleneck in my work, and I want the preparation to help me through it.
A brief that helps me decide.
Confirm the pilot’s success criterion before building the demo.
Lead with the decision I need to make.
The champion wants faster onboarding. The technical lead asked how we would measure it. The current pilot outline names features, but no agreed target.
Give me the evidence that changes the recommendation.
Customer call: stated goal.
Technical follow-up: measurement question.
Pilot outline: target still missing.
In the working version, each reference opens the original source.
Who owns the baseline, and what improvement would justify expanding?
Keep uncertainty visible.
“Before we tailor the demo, can we agree what a successful pilot would change for your team?”
Offer something I can adapt and use.
Not everything needs AI. And when I do reach for it, I have to budget an unexpectedly large amount of time to get from “that looks pretty good” to “this is good enough to submit where it matters.” That’s the right side of the U, and it’s where I spend most of my day now. And refinement continues after delivery: real use reveals what needs to change, and those lessons shape the next round of planning. That ongoing learning loop is what I take from Brooke’s argument.
Nobody has the new pattern figured out yet
Who has the time to read, think, edit, and act on all of those outputs? Even if we accept that quality goes up with quantity, who is accountable for the impact of that net new workload? When productivity rises, how should the gains be shared between the business and the people doing the work? Do we just expect people to work harder for the same pay?In his 2026 survey of tech workers (opens in a new tab), Noam Segal found that losing a job to AI ranked second to last among the concerns people named. What rose to the top was “the expectation to do more for the same pay.” Burnout above moderate rose from 44.7 to 54.7 percent between his 2025 and 2026 samples. Self-reported, and the samples aren’t matched, but it lines up with what I hear in every room I’m in. How much more can a person reasonably take on to “keep up with AI,” when keeping up is impossible?
I don’t have clean answers. What I have is a growing suspicion that the exhaustion isn’t a personal failing or a phase. It’s structural. Organizations are built around a narrow expected range of human output, and when one person’s output triples, the approvals, staffing, and coordination around them don’t know what to do with it.Ethan Mollick made this argument in July 2026 (opens in a new tab): “organizations are built around a narrow expected range of human productivity in a role,” and the “approvals, staffing models & coordination systems cannot absorb” much more. It’s an authored argument rather than a study, and it helps explain why individual gains may not yet show up in organizational outcomes. The gain in capacity is real. What happens to it is a decision nobody has made yet.The Enterprise AI Playbook, chapter 6 (opens in a new tab) records different staffing destinations among its selected successful deployments. It does not settle which choice a business should make or predict a worker’s job security. The distinction matters: an observed gain in capacity and a decision about staffing are separate claims.
Parallel work makes it worse before it makes it better. I can keep several independent workstreams moving overnight, and then someone has to understand all of the results and act on them in the morning. That someone is me.Katie Parrott’s account of running three models in parallel (opens in a new tab) treats attention as the constraint. That resonates with the coordination work I want to make visible here. It’s a personal practitioner account, not an organizational productivity estimate: parallel production still needs someone to integrate the results.
The clearest way I can say it: there are three jobs nobody had two years ago, and they’ve been stacked on top of the job everyone already had.
- Learning AI. Building the intuition for what these tools are natively good at, where they’re jagged, and how that changes every month.
- Managing AI. Planning, refining, reviewing, and deciding when an agent’s work is good enough to leave your hands. The two big ends of the U.
- Managing people learning AI. Teaching, setting standards, judging output you didn’t produce, and deciding whose week changes and how.
This is why so many shiny pilots evaporate. Someone set them up and nobody planned for the three jobs that follow the setup. That’s the felt sense of exhaustion. New work nobody named, stacked on old work nobody removed.
Test the new shapes on purpose
If nobody has the pattern, the honest move is to go find it, deliberately and in small pieces. I’ve stopped trying to design the perfect workflow up front. I run experiments sized so that failing costs almost nothing and succeeding teaches something.
Starting small makes everything tangible. People are experts in their own daily work, so they can intuitively run fast feedback loops for a tool that makes that work faster, easier, and better. When it does, they naturally want more. And starting small lowers the stakes in a way that makes it okay to fail. If I botch something as trivial as prepping for a meeting, I’m no worse off than if I hadn’t prepared at all (my default). If the meeting prep skillz are even slightly useful, I’ve gotten some value and I can feel the jagged frontier of what AI is natively good at and where it struggles.
To keep the initial momentum, I build reinforcement loops around the experiments. Things as simple as a weekly lunch-and-learn or an #ai-learnings channel where people share one thing that worked can do wonders. As people gain confidence, they get ambitious on their own. You’ll notice them reaching for workflows at higher levels of complexity and abstraction. Creating fertile soil for new AI experts to grow is one of the hardest and most rewarding parts of this work.
One rule I hold myself to: measure before I believe. A faster task can leave the real constraint untouched, and a smaller queue is not always a solved problem.The specifics of baselines, attribution, deflection versus resolution, and what to count before calling something a financial return live in the field guide, under Follow the result into someone’s day. This essay is about why that discipline is a survival skill, not a reporting chore.
Make room to learn
Here is the part I most want leaders to hear. Learning adds work before it returns capacity. A new workflow costs more than it saves at first, because people still do the work and now also have to check it closely enough to know whether it’s right. In my own projects, my rule of thumb is about a third more effort per workflow run while the new approach settles in, compared with running the previous workflow. It depends on the people, the complexity of the work, and how quickly they learn to review. What nobody reports is zero.
Enthusiasm doesn’t create hours in the day. Leaders decide where the hours come from: changing priorities, adding support, accepting a bounded stretch, or some other workable arrangement. What doesn’t work is pretending the learning is free, or extracurricular, or something people will absorb on top of everything else because the tools are exciting.Natalia Quintero’s executive guide (opens in a new tab) is blunt about this: champions need protected time, at least two days a month in Every’s experience, and a clear mandate, or the AI work gets squeezed out by the day job. She also names the cost of going from a workflow that’s 60 percent right to one that’s 95 percent right: examples, evaluation, feedback, human review, and maintenance. That’s her consulting heuristic, not a researched minimum, and it matches what I see.
Once capacity does appear, that’s a second decision, and it belongs in the open. It can go toward deeper customer relationships, neglected quality work, new experiments, or a cost objective. I want that choice connected to business needs and discussed candidly with the people whose work is changing, because they’ll notice either way. That conversation should include how they share in the gains: compensation, opportunities to grow, or breathing room in a demanding day.
And I want the incentives to support the method. Recognition for a flashy demo is easy. Teaching colleagues, maintaining shared standards, and reporting a disappointing result deserve support too. If the only thing that gets rewarded is the first demo, the three new jobs never get done, and the exhaustion compounds.
Some of the room has to be made by individuals, too. When the work has doubled and the pay hasn’t, that’s a conversation to have with your manager early and on purpose, before the squeeze becomes the job.
Learning AI, learning to manage AI, and managing people learning AI are different and new responsibilities, and most leaders are doing all three on top of what they were already doing (plus the extra work from AI efficiencies). Oy! Better we say that plainly than pretend the new patterns of work are settled.
Thriving with AI @Work isn’t keeping up. Keeping up is impossible. It’s choosing what to learn, testing it in small enough pieces to learn fast, and making real room for it, together, in this moment of flux.
Read: Build the work people can actually use · Compare notes with me