What AI Can't Grade: Skills That Still Matter

On 2 July 2026, Phil Chen published a post about careers in an AI-native world. He is a founder whose career has included work at Scale AI, DeepMind and OpenAI. The line that stayed with me was this:

"AI models get better at anything you can write a loss function for, and school is mostly loss functions."

A loss function is simply a way to score an output. Is the answer correct? Is the report complete? Does the spreadsheet calculate the right result?

Chen's argument is that, when a task can be scored clearly, AI systems will keep getting better at doing it. He says the valuable work of the next decade is what cannot be graded within the span of model training. For a business, that does not mean every scored task disappears. It means the human advantage moves toward the work that cannot be reduced to a clean score.

For a business owner, this is not an abstract career debate. It is a useful way to decide where AI belongs in the company, and where it does not.

Most knowledge work is gradable work

Much of what we call knowledge work has an implicit grading rule.

A report is judged on clarity, accuracy, and completeness.

A spreadsheet model is judged on whether the numbers work.

A contract draft is judged on whether it covers the relevant clauses.

These are useful tasks. They are also tasks where AI can already produce a reasonable first result.

The mistake is to treat that first result as the whole job. A draft still needs context. A model still needs someone to question its assumptions. A contract still needs qualified review. The more consequential the work, the less a generic score tells you.

Chen is not saying that humans become irrelevant. Across his post, the scarce work is choosing the right problem, deciding what good looks like, and getting the last part right.

1. Find the right problem

A well-defined problem is easier to solve than a poorly chosen one.

That distinction matters in a real company. A team can spend weeks improving a report while the real bottleneck sits in the handoff before the report, or in the decision nobody owns after it is delivered.

Problem selection means walking into an unfamiliar workflow and asking:

• Where does work actually stop or slow down?
• Which decision keeps coming back?
• What information is missing when that decision is made?
• What would change if this one part became clearer or faster?

That is not mainly a tool question. It is an operating question.

When I work on an AI OS Blueprint, the valuable design work is not making AI sound clever. It is mapping where leverage could exist, checking the actual environment, and deciding whether a bounded first installation is worth doing at all. Sometimes the right answer is AI. Sometimes it is a simpler process change. Sometimes the answer is to defer.

2. Finish the last mile

In a startup context, Chen quotes Alfred Lin's idea that the last 10% of a project can contain 90% of the work and 90% of the reward.

AI can produce a decent draft, prototype, or analysis from a weak prompt. That is useful. It is not the same as a system people can rely on every week.

The last mile is where the awkward questions appear:

• What happens when the input is incomplete?
• Who checks the output?
• What is the escalation path?
• Can someone stop the workflow?
• Is the process documented well enough for another person to run it?
• Will the team actually use it after the demo?

This is why a 30-second AI demo proves very little. The gap between a working prototype and a dependable workflow is not a footnote. It is the work.

3. Build trust that compounds

Chen's post also points to time, relationships, and reputation as resources that remain scarce. Tools become easier to access. Trust does not.

That matters inside a business. Employees need to know when an AI output can be used and when it must be checked. Customers need to know who stands behind a message or decision. Owners need to know who can stop a workflow when it behaves badly.

The tool is not the trust. The operating rules around the tool are what make its use credible.

A company that uses AI well is not necessarily the company with the most tools. It is the company that can explain what each tool does, where human judgment remains, and who owns the next decision.

What this means for your business

You do not need to become an AI researcher. In his fifth point, Chen says he does not believe artificial superintelligence will replace all humans in knowledge work jobs, because humans have different capabilities in selecting meaningful problems for it to solve and in allocating capital to those problems. His sixth point is that being a researcher is a mentality, not an occupation. In a small or mid-sized company, that mentality looks practical:

1. Pick one recurring bottleneck.
2. Describe what a good result looks like.
3. Separate the repeatable work from the judgment-heavy work.
4. Test one bounded intervention.
5. Review the result with the people who have to live with it.

That description of "good" is your working score. It gives an AI-assisted workflow something to aim at, and it gives a human a clear point at which to reject the output.

The businesses that benefit from AI will not simply be the ones with the best model. They will be the ones that decide clearly what should stay human, what should be AI-assisted, and what can be delegated safely.

That is a leadership decision before it is a technology decision.

The real question

For a particular person or company, the answer depends on practice, not optimism.

Can you identify the problem worth solving? Can you define what good looks like? Can you finish the last mile? Can people trust the result?

Those are not easily gradable skills. They are still business skills.

Source

Phil Chen, "Career advice in the age of AI," X, 2 July 2026: https://x.com/philhchen/status/2072793818945167475