When a model can finish the task in minutes, what is left that someone will pay for? Austin Marchese, who ran an engineering team at J.P. Morgan and now advises companies on using AI, argues what people pay for is moving to four places: choosing the task, doing things others do not know AI can do, feeding your results back into your tools, and proving you can deliver. One yardstick runs through all four. Value is what someone gets, divided by what it costs them in time, effort, and money.
01Deciding is now the hard part
The hard part of a job used to be doing the thing. Now it is knowing which thing to do. Marchese manages four engineers on a sports app with over 100,000 users, and what he is paid for is helping decide what gets built and how, not writing the code.
“people are 10 times more productive in the other 10%”
Dario Amodei, Anthropic CEO, “Now That Claude Does Everything, Here's What AI Can't Replace” · 1:26Producing options got cheap; choosing did not. A model will hand you a hundred variations of a page, but you can put only one in front of your boss, your customers, or your audience, and you get a limited number of those chances. As building speeds up, a good pick starts paying off sooner, and a bad one gets copied into more work before anyone catches it.
Three habits raise the quality of your decisions:
- Sort decisions by reversibility. Jeff Bezos calls these one-way and two-way doors. A two-way door can be walked back, so decide fast and fix it later if needed. A one-way door cannot, so think hard first. Asking “can I undo this?” before deciding tells you how much deliberation each one deserves.
- Rehearse the audience before they see it. Have the model interview you about the person on the receiving end of the work, a boss or a customer: what they care about, what they reject. Then have it play that person and critique your draft. You get the feedback before the real person does, while changes are still cheap.
- Ask one more question than the instructions did. Told to write an email for a product launch? Executing exactly that is what a model does, faster and cheaper. Asking who the audience is, noticing it splits into two groups, and writing one email for each produces a better result for one extra question. The failure on the far side is real too: 45 minutes debating whether the email opens with “hey” or “hi” is deliberation spent where it changes nothing. One question deeper, then ship.
Executing instructions word for word is how average work gets made: it is the most likely answer to the request as written. Naval Ravikant’s line for that, quoted in the video (4:14), is “there is no demand for average”. Why models produce that same average by default is the subject of Why AI makes slop and how to catch it.
02Give away what is easy for you
AI capability grows faster than people learn what it can do. Marchese shows a text from a friend, astonished at how easy it was to build a working website by describing it to a model, a thing anyone reading this guide has known for years. That gap is an opening: tasks that are routine for you are still astonishing to most people.
The move is to find a task that is quick for you and valuable to one specific person, then do it without being asked. An employee can build the small internal tool their manager never requested but clearly needs. A business can refresh a client’s dated website for free.
The usual objection is that if a model did most of it, the work should not count. Run the value equation from the receiver’s side. They get the outcome, it cost them nothing, and how easy it was for you never enters their math.
“AI is the tool, and the output is what matters”
Austin Marchese, “Now That Claude Does Everything, Here's What AI Can't Replace” · 8:54The gap will close as people catch up, so the point of the free work is what outlasts it: a relationship with someone who now knows firsthand that you deliver. Section 04 is where that becomes the payoff.
03Feed your results back in
Everyone rents the same models, so the edge comes from what you feed back into them. A memo from Microsoft CEO Satya Nadella, read in the video, splits a firm’s assets into the judgment of its people and the AI systems the firm builds and owns, and argues the second is worthless without the first.
“Without human direction, you have compute running in circles.”
Satya Nadella, Microsoft CEO, “Now That Claude Does Everything, Here's What AI Can't Replace” · 9:47Marchese’s concrete case: an engineering firm he works with produces a building-emissions report that New York City requires under a law called Local Law 97. They built a Claude skill to generate it. The first ten to twenty runs hit edge cases, and after each one the firm’s engineers corrected the skill with feedback only they could give. Now it produces the report the way they want it, every time. No competitor can copy that result by picking the same model, because the corrections came from private experience.
The loop has three steps:
- Pick something you do repeatedly. A one-off task gives you nothing to improve against; repetition is what turns feedback into an asset.
- Make it a skill, not a one-off prompt. A skill is a saved, reusable set of instructions the model follows every time. Expertise you can install covers what skills are; Capture the model’s discipline covers keeping a top model’s way of working when the model itself changes.
- Fold each run’s mistakes back into the skill. Every run that goes wrong is a correction only you can supply. Turn repeated fixes into rules is the working procedure.
Marchese flags this shift as the time-sensitive one: every repeat of the task you run without capturing the lesson is a correction lost.
04Prove it before they ask
Polished output is now suspect. Recruiters report they often cannot tell a strong candidate from a weak one who used AI to polish a resume, so many have stopped trusting the inbox: they source people directly and look at portfolios, public code, and paid trial projects instead. The same doubt now lands on anything you claim you can do.
Proof comes in four forms. Each is harder to produce than the last, and harder to fake:
| Proof | What it is |
|---|---|
| Tell them | A resume, a portfolio, a case study: your account of what you can do |
| Show them | The actual task, done: a working sample, the exact outcome they want |
| Someone vouches | Another person stakes their credibility on you |
| They already got value | This person has firsthand results from you |
The last row is why section 02 matters beyond the quick win. Every unasked-for delivery moves one more person into the bottom row, and that record compounds while the gap that made the work easy closes behind you.
Further reading
- Austin Marchese · Now That Claude Does Everything, Here’s What AI Can’t Replace, the video this guide synthesizes