Episode 01
AI Did in 5 Minutes What Took You 15 Years to Learn. Now What?
What experience is actually made of, and which parts of it a tool can reproduce.
The video for this episode is not yet published.
The written companion, framework, worksheet, and sources for “AI Did in 5 Minutes What Took You 15 Years to Learn. Now What?” are available below. The video will appear here when it goes live on YouTube.
The question beneath the video
If AI makes part of my expertise cheaper, what exactly did all those years leave behind?
The distinction
- 01Output
- 02Workflow segment
- 03Judgment and responsibility
- 04Market price
Written companion
This is the written companion to Episode 1. It is not a transcript. It goes slower than the video, shows more of the reasoning, and points to the sources so you can check them yourself.
The moment
It usually happens in an ordinary meeting. Someone younger, or simply more curious, pastes a request into a tool and reads out the result. The result is not perfect. It is also not bad. It resembles what you would have produced on a first pass, and it took about as long as it takes to make coffee.
What follows is not usually panic. It is quieter than that. It is a small recalculation of your position that happens before you have decided to make it. You keep contributing to the meeting. Some part of you has started doing arithmetic about fifteen years.
This episode takes that arithmetic seriously. It does not tell you the tool is worse than it looked. It does not tell you your job is safe. It asks a more precise question: what, exactly, did the years build, and how much of that did the tool just reproduce?
Four things that get counted as one
When people say my expertise, they are pointing at a bundle. The bundle has at least four parts, and they have different relationships to what a tool can do.
- 01Output
- 02Workflow segment
- 03Judgment and responsibility
- 04Market price
Output is the artifact. The draft, the analysis, the design, the code, the plan. It is the visible thing, and it is what the tool produced in the meeting.
The workflow segment is the stretch of the process where that output lives. Producing a draft is one segment. Knowing what to ask for, checking what came back, deciding whether it can go out, and answering for it afterward are other segments. The tool reached into one segment, and it did so impressively.
Judgment and responsibility is the part that does not show up in the artifact. It is knowing when the output is wrong in a way that will cost someone. It is the timing of a question. It is the exception you remember because you were there when it went badly. And it is the plain fact that when the draft goes out under a name, someone answers for it, and the tool does not.
Market price is what someone was willing to pay for the bundle. It was never a measurement of the years. It was a measurement of scarcity, and scarcity in the first segment is changing fast.
- Output: the artifact
- Workflow segment: where the artifact lives
- Judgment and responsibility: what the years built
- Market price: what someone paid for the bundle
The pain in the meeting comes from treating these four as one thing and watching a tool reproduce it. Separated, the picture is different. The tool reproduced the output and a piece of a segment. The judgment is what you used, half a second later, to notice what the draft got wrong. The price is being renegotiated, which is real, and which is a different problem from the years having been wasted.
What the evidence supports, and what it does not
It would be easy to build this argument on reassurance. Voyagers Beyond tries not to do that. Here is what the research actually shows, with the boundaries drawn.
Working paper
Generative AI at Work
Erik Brynjolfsson, Danielle Li, Lindsey Raymond · National Bureau of Economic Research · NBER Working Paper 31161; later published in The Quarterly Journal of Economics · 2023-04
What the source supports
- In this setting, a generative AI assistant increased average productivity, measured as issues resolved per hour.
- The gains were concentrated among novice and lower-skilled agents. Experienced, high-skill agents saw minimal improvement.
- The authors interpret the tool as spreading the practices of more able workers to newer ones.
What it does not support
- That experienced workers in other fields will see the same pattern. The study covers one company and one kind of work.
- That AI assistance makes experience worthless. The tool learned from the experienced agents' own behavior.
- Any claim about job loss, wages, or long-term careers. The study measures task performance, not employment outcomes.
What Sougata is doing with it
Sougata uses this as evidence that the visible output gap between novice and expert can shrink quickly, while noting that the expert's judgment is what the tool was trained to imitate. That distinction between output and judgment is central to Episode 1 and is Sougata's synthesis, not the paper's claim.
Working paper
Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality
Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, Karim Lakhani · Harvard Business School · HBS Working Paper 24-013; forthcoming in Organization Science · 2023-09
What the source supports
- AI capability is uneven across tasks that look similar in difficulty. The authors call this the jagged frontier.
- For tasks inside the frontier, consultants with the tool performed measurably better on speed and rated quality.
- For a task outside the frontier, tool use was associated with worse accuracy.
What it does not support
- That the frontier is fixed. The authors describe it as moving and hard to see from the inside.
- That professionals can reliably tell in advance which side of the frontier a task sits on.
- Any general claim about consulting careers, hiring, or the value of experience over time.
What Sougata is doing with it
Sougata uses this to support one specific point: the segment of a workflow that a tool handles well is not the whole workflow, and knowing where the tool fails is itself a form of experienced judgment. The framing of workflow segment versus judgment is Sougata's, not the paper's.
Journal article
Experimental evidence on the productivity effects of generative artificial intelligence
Shakked Noy, Whitney Zhang · Massachusetts Institute of Technology · Science, vol. 381, issue 6654 · 2023-07
What the source supports
- On short professional writing tasks, access to the tool reduced average time and raised average rated quality.
- The gap between lower- and higher-performing participants narrowed with the tool.
- Participants' reported concern and excitement about AI both rose after exposure.
What it does not support
- That the same effects apply to long, high-stakes, or domain-specific work. The tasks were short and general.
- That quality as judged by evaluators equals the quality an experienced professional would be accountable for.
- Anything about what happens to careers, pay, or demand for writers over time.
What Sougata is doing with it
Sougata uses this to describe how a first draft, the visible output, became cheap for a whole class of tasks. The claim that responsibility for the draft did not become cheap is Sougata's interpretation, and the paper does not test it.
Read together, these studies support two things and refuse to support a third. They support the claim that visible output became cheap for a wide class of tasks, and that the gap between novice and expert output narrowed. They also support the claim that the tool's competence is uneven and hard to see from the inside, so that knowing where it fails is itself a form of experience. What they do not support is any prediction about your career. They measured tasks over weeks. They did not measure lives over decades.
The part that is hard to say
There is an old observation, from the philosopher Michael Polanyi, that people know more than they can put into words. A skilled person cannot fully write down what they are doing when they do it well. That is not mysticism. It is a description of how expertise actually sits in a person: as noticing, as timing, as a feel for when something is off.
Book
The Tacit Dimension
Michael Polanyi · University of Chicago Press (originally Doubleday, 1966) · 1966
What the source supports
- The claim that skilled practice includes knowledge the practitioner cannot fully articulate.
- The observation that explicit rules never capture the whole of a skill.
What it does not support
- Any claim about artificial intelligence, which the book predates by decades in its modern form.
- The claim that tacit knowledge cannot be imitated in output. Polanyi is describing knowing, not the reproducibility of results.
What Sougata is doing with it
Sougata uses Polanyi as a vocabulary for the part of experience that shows up as noticing, timing, and knowing when a result is wrong. Applying that vocabulary to AI-assisted work is Sougata's extension and should be read as interpretation.
This creates a specific problem for experienced people right now. The part of their expertise that is hardest to reproduce is also the part that is hardest to show. Output is easy to show. Judgment is not. So when a tool reproduces the output, the experienced person looks, from the outside, as if they have been matched. From the inside, they know they have not been. And they have no easy way to demonstrate the difference.
That is the problem the Second Door Plan is built for.
Replaceable, irreplaceable, and a third option
Most advice for this moment tells you to become irreplaceable. Learn the tool faster than everyone else. Move up the value chain. Specialize until no one can touch you.
That advice has a hidden assumption: that there is one room, one role, one buyer for your value, and your job is to defend your position in it. The strategy is to make yourself so valuable in that one place that you cannot be removed from it. It is an exhausting strategy, and it is fragile, because it depends on a place you do not control.
The opposite of being replaceable is not becoming irreplaceable. It is having more than one place where your value can live.
A second door is not a second job. It is a second place where a specific piece of your judgment is worth something. It might be a different team, a different industry, a different kind of organization, or a different form: teaching, reviewing, advising, writing. What matters is that it exists, that you have tested it, and that you know its name.
The Second Door Plan
The method has five stages. The framework page goes through each in detail, with an illustration and a worksheet. Here is the shape.
- Find. Locate one recurring problem your experience taught you to see before others do. A problem, not a skill.
- Verify. Check with two or three people who do adjacent work that the problem is real, recurring, and costly to someone.
- Extract. Write the judgment down. The signals, the order, what wrong looks like, the exceptions. One page a stranger could follow.
- Demonstrate. Apply the page to a real or clearly labeled case in a form someone else can inspect.
- Test. Put the demonstration in front of an audience outside your current role and see whether it produces a signal.
Now what
The question in the title is not rhetorical. Here is a direct answer. The fifteen years built an output you can produce, a segment of a workflow you know well, a body of judgment that mostly lives in your noticing, and a price that was attached to the bundle. A tool reproduced the first and part of the second. The third is intact, and it is the part that is hard to show. The fourth is changing.
The work in front of you is not to defend one room. It is to make the third thing visible, and then find out where else it matters. Start with the worksheet. It takes about ninety minutes, and it ends with one action and a date.
We do not know what happens next. We can still examine what is happening now.
Use
The framework
Try
The worksheet
Worksheet v1.0
The Second Door Plan Worksheet
To take one recurring problem your experience taught you to notice and work it through the five stages of the Second Door Plan, ending with a concrete test you can run within a few weeks.
Educational reflection material. Not psychological, career, financial, legal, or medical advice, and not a validated instrument. No particular result is promised. Read the full scope statement.
Verify
Evidence and sources
Each source is separated into what it supports, what it does not, and what Sougata is doing with it.
Working paper
Generative AI at Work
Erik Brynjolfsson, Danielle Li, Lindsey Raymond · National Bureau of Economic Research · NBER Working Paper 31161; later published in The Quarterly Journal of Economics · 2023-04
A field study of a generative AI assistant rolled out to several thousand customer support agents. The tool raised issues resolved per hour, with the largest gains among newer and lower-skilled agents and small effects for the most experienced.
What the source supports
- In this setting, a generative AI assistant increased average productivity, measured as issues resolved per hour.
- The gains were concentrated among novice and lower-skilled agents. Experienced, high-skill agents saw minimal improvement.
- The authors interpret the tool as spreading the practices of more able workers to newer ones.
What it does not support
- That experienced workers in other fields will see the same pattern. The study covers one company and one kind of work.
- That AI assistance makes experience worthless. The tool learned from the experienced agents' own behavior.
- Any claim about job loss, wages, or long-term careers. The study measures task performance, not employment outcomes.
What Sougata is doing with it
Sougata uses this as evidence that the visible output gap between novice and expert can shrink quickly, while noting that the expert's judgment is what the tool was trained to imitate. That distinction between output and judgment is central to Episode 1 and is Sougata's synthesis, not the paper's claim.
Working paper
Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality
Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, Karim Lakhani · Harvard Business School · HBS Working Paper 24-013; forthcoming in Organization Science · 2023-09
A preregistered field experiment with several hundred management consultants. On tasks inside the model's capability, consultants using GPT-4 finished more tasks, faster, and at higher rated quality. On a task designed to sit outside that capability, those using the tool were more likely to get the answer wrong.
What the source supports
- AI capability is uneven across tasks that look similar in difficulty. The authors call this the jagged frontier.
- For tasks inside the frontier, consultants with the tool performed measurably better on speed and rated quality.
- For a task outside the frontier, tool use was associated with worse accuracy.
What it does not support
- That the frontier is fixed. The authors describe it as moving and hard to see from the inside.
- That professionals can reliably tell in advance which side of the frontier a task sits on.
- Any general claim about consulting careers, hiring, or the value of experience over time.
What Sougata is doing with it
Sougata uses this to support one specific point: the segment of a workflow that a tool handles well is not the whole workflow, and knowing where the tool fails is itself a form of experienced judgment. The framing of workflow segment versus judgment is Sougata's, not the paper's.
Journal article
Experimental evidence on the productivity effects of generative artificial intelligence
Shakked Noy, Whitney Zhang · Massachusetts Institute of Technology · Science, vol. 381, issue 6654 · 2023-07
A preregistered online experiment in which college-educated professionals completed mid-level writing tasks with or without access to ChatGPT. Access reduced time taken and raised rated quality, and narrowed the gap between weaker and stronger writers.
What the source supports
- On short professional writing tasks, access to the tool reduced average time and raised average rated quality.
- The gap between lower- and higher-performing participants narrowed with the tool.
- Participants' reported concern and excitement about AI both rose after exposure.
What it does not support
- That the same effects apply to long, high-stakes, or domain-specific work. The tasks were short and general.
- That quality as judged by evaluators equals the quality an experienced professional would be accountable for.
- Anything about what happens to careers, pay, or demand for writers over time.
What Sougata is doing with it
Sougata uses this to describe how a first draft, the visible output, became cheap for a whole class of tasks. The claim that responsibility for the draft did not become cheap is Sougata's interpretation, and the paper does not test it.
Book
The Tacit Dimension
Michael Polanyi · University of Chicago Press (originally Doubleday, 1966) · 1966
A short philosophical work arguing that a great deal of human knowledge cannot be fully put into words: people know more than they are able to articulate, and skilled practice depends on that unstated knowledge.
What the source supports
- The claim that skilled practice includes knowledge the practitioner cannot fully articulate.
- The observation that explicit rules never capture the whole of a skill.
What it does not support
- Any claim about artificial intelligence, which the book predates by decades in its modern form.
- The claim that tacit knowledge cannot be imitated in output. Polanyi is describing knowing, not the reproducibility of results.
What Sougata is doing with it
Sougata uses Polanyi as a vocabulary for the part of experience that shows up as noticing, timing, and knowing when a result is wrong. Applying that vocabulary to AI-assisted work is Sougata's extension and should be read as interpretation.
Continue the question
Related questions
Work & AI
What were my fifteen years for?
A tool produced in minutes something that resembles work you spent years learning to do. The question is not rhetorical. It deserves a real answer about what those years actually built.
Identity & Worth
Why do I feel replaceable when I still have my job?
Nothing has happened yet. The role, the pay, and the team are all still there. The feeling arrived anyway. That gap between the event and the feeling is worth examining on its own.
Uncertainty & Agency
How do I adapt when the target keeps moving?
Every piece of advice says adapt. Almost none of it says to what. When the thing you would adapt to changes every few months, adaptation as a strategy needs a different definition.
Continue