What cannot be replaced is also what cannot be taught. Naval called it specific knowledge. AI is clarifying which knowledge actually qualifies — and the answer is narrower and more valuable than most people expect.
Naval Ravikant defined specific knowledge as the knowledge that cannot be trained. "It is found by pursuing your genuine curiosity and passion rather than whatever is hot right now. It looks like play to you but work to everyone else." (per nav.al, 2018)
The AI age is making this distinction more important, not less. AI systems are very good at generalizable knowledge — the kind that can be described in a training set, systematized into a process, or expressed as a pattern in historical data. This includes most of what most knowledge workers do for most of their workday.
What AI is not good at — yet, and possibly never, depending on your view — is the judgment that comes from irreplaceable experience, the pattern recognition that comes from having made real decisions with real consequences, and the trust that comes from demonstrated skin in the game over years of visible performance. These are the components of specific knowledge. They are also the components that define value in a post-AI economy.
The question worth sitting with is not "will AI replace my job?" It is "which part of what I do is generalizable knowledge, and which part is specific knowledge?" The first part will be replaced. The second part will become more valuable as the first part gets cheaper.
The tasks that AI replaces first are not the simple ones. They are the systematic ones.
A task is easy to automate if it can be described in a sufficiently detailed prompt. Legal contract review — systematic. Financial report summarization — systematic. Code that follows documented patterns — systematic. Proposal writing that follows a template — systematic. Market research that aggregates published data — systematic.
None of these require novel judgment. They require the application of known frameworks to new inputs. That is precisely what large language models are designed to do.
Anthropic's research on AI capabilities as of 2025 indicates that models perform best on tasks with clear evaluation criteria and sufficient training data — which describes the majority of structured knowledge work (per Anthropic, 2025). The ceiling on AI performance in structured tasks has not been reached.
The relevant implication for agency owners and knowledge workers is not dystopian. It is organizational. The generalizable tasks that currently occupy 60 to 70 percent of senior knowledge-worker time in most service businesses can be delegated to AI systems — not perfectly, but well enough to free that time for the judgment work that remains genuinely human.
The agency that uses AI to handle systematic work and redeploys senior time to judgment work is not threatened by AI. It is leveraged by it.
Not all non-AI work is specific knowledge. Some tasks are simply not yet automated — due to cost, access, or organizational inertia — but will be. Specific knowledge is a narrower category.
The three clusters that consistently resist full AI replacement:
Judgment in genuinely ambiguous situations. When the right answer is not derivable from historical patterns because the situation is structurally novel, human judgment is required. This is not common in routine knowledge work but is the defining feature of the most valuable strategic decisions. Which client to turn down. Which product bet to place. How to handle a relationship crisis with no precedent.
Pattern recognition from irreplaceable experience. Some knowledge is encoded in physical and social experience that cannot be compressed into training data. A founder who has built and sold two companies has pattern recognition about organizational dysfunction that cannot be transferred by reading about organizational dysfunction. The experience of having made decisions with real consequences, over years, under genuine uncertainty, creates a form of judgment that is not available from any other source.
Trust earned through demonstrated skin in the game. Clients pay premium rates not for information — information is increasingly free — but for the confidence that the person advising them has staked their own reputation on their recommendations. This is the "skin in the game" signal that Nassim Taleb identified as the foundation of genuine expertise: you are trusted because you have been wrong before, have paid the cost of being wrong, and have improved because of it (per First Round Review, 2024).
All three of these are, in different ways, irreplaceable by AI systems — not because AI lacks intelligence, but because they depend on accumulated, situated, consequence-tested experience that no training process can replicate.
For agency owners and practitioners, the question is concrete: which parts of the work belong to which category?
The pattern is clear. AI handles the systematic, the generalizable, the well-structured. Humans own the judgment, the relationship, the novel, and the consequential.
For most agencies, this means the work that should occupy the most senior time is not the work that currently does. The 70 percent of senior-staff time spent on systematic tasks — drafting, formatting, researching, reporting — can be progressively automated. The 30 percent spent on genuine strategic judgment, client trust, and novel problem-solving is the work that cannot be compressed.
Specific knowledge is identified by three tests.
First: would someone pay meaningfully more for your judgment specifically than for a competent generic answer? If the answer is yes, you have some specific knowledge in that domain. The premium over the generic reflects the market's assessment of the irreplaceable component.
Second: did you develop this knowledge through genuine curiosity and consequence — not through a curriculum? If you could have learned it the same way from a well-structured course, it is not specific knowledge. If the learning required you to make real decisions with real stakes, it is.
Third: is it something that feels like play to you and looks like work to others? Naval's original framing of this test is precise. The work that is genuinely energizing, that you would do without being paid, and that most people find effortful — that is the work closest to your specific knowledge.
Protecting specific knowledge in the AI age is about investment, not resistance. The agency owner who accelerates AI adoption in all the generalizable work creates more time for the specific-knowledge work that the AI adoption cannot touch. This is the correct response to AI: not to resist it in the systematic tasks it is suited for, but to use it aggressively so that senior human time is reserved for the work that compounds.
See striveloom.com/services to see how we have restructured our service delivery around this distinction — AI and automation on the systematic side, senior judgment on the specific-knowledge side.
The agencies that will compound their value through the AI transition are not the ones that resist AI adoption. They are the ones that adopt AI in all the systematic work and redeploy the freed capacity into the specific-knowledge work that becomes more scarce and more valuable as the systematic work becomes cheaper.
This is not a paradox. It is the same compounding logic that applies to every form of leverage. The resource that becomes more available — in this case, AI-executed systematic work — gets cheaper. The resource that becomes more scarce — irreplaceable human judgment, earned trust, novel pattern recognition — gets more expensive.
Invest in your specific knowledge. That is the part of your business that compounds through the AI transition. Everything else is subject to the deflationary economics of a technology that improves faster than any individual can.
The long game belongs to the people who figured out which game they are actually playing.
Naval defines specific knowledge as the knowledge that cannot be trained or taught through a curriculum. It is found through pursuing genuine curiosity and passion rather than following conventional career paths. It feels like play to the person who has it and looks like work to everyone else. Because it is not teachable in a structured sense, it cannot be commoditized or easily replicated. In the AI age, this definition matters more than ever because the knowledge that can be described and systematized — and therefore taught — is precisely the knowledge that AI can learn from a training process.
The AI-resistant skills cluster around three areas: judgment in genuinely ambiguous situations where historical patterns do not apply, pattern recognition from irreplaceable accumulated experience with real consequences, and trust earned through demonstrated skin in the game over time. These share a common trait: they depend on situated experience that cannot be compressed into training data. Relationship-dependent advisory work, novel strategic decisions with organizational stakes, and expertise earned through failure and recovery are the clearest examples in a professional services context.
The correct response is to accelerate AI adoption in systematic tasks and redeploy the freed senior time into specific-knowledge work. Agencies that use AI to handle research, first-draft writing, reporting, and structured analysis can redirect senior hours to genuine strategic judgment, client relationship management, and novel problem-solving — work that becomes more valuable precisely because it is scarcer. Resistance to AI adoption is not a strategy for preserving specific knowledge. It is a strategy for remaining busy with generalizable work while competitors free their senior time for higher-value output.
The current honest answer is: not fully, and not for the specific knowledge that matters most. AI systems learn from training data. Specific knowledge, as Naval defines it, is developed through consequence-tested experience that is not available in any dataset. The judgment of a founder who has built and sold companies, the trust of an advisor who has been wrong and has paid for it, the pattern recognition of a practitioner who has navigated novel situations for twenty years — none of this is in a training set. AI can approximate generalizable outputs in these domains. It cannot replicate the situated judgment behind the best decisions.
Three tests help: First, would people pay meaningfully more for your specific judgment than for a competent generic answer? The premium reflects the market's assessment of the irreplaceable component. Second, did you develop this knowledge through genuine curiosity and real consequences, not through a curriculum you could replicate? If the learning required real decisions with real stakes, it is more likely specific. Third, does it feel like play to you but work to others? Naval's test is precise — the work that is energizing for you specifically, that you would do without pay, and that most people find effortful, is the work closest to your specific knowledge.
Generalist knowledge workers whose primary output is systematic work — research, structured writing, data analysis, template-based deliverables — will face direct price competition from AI-assisted workflows. The response is specialization toward specific knowledge: developing deeper expertise in domains where judgment and experience matter, accumulating the kind of track record that creates trust, and positioning around the work that AI cannot systematize. The generalist who pivots to 'AI-assisted generalist' is a temporary position. The specialist who uses AI to amplify specific knowledge is a durable one.
Founder & CEO of Striveloom. Software engineer and Harvard graduate student researching software engineering, e-commerce platforms, and customer experience. Builds the agency that ships like software — one team, one pipeline, one platform. Writes on AI agencies, web development, paid advertising, and conversion optimization.
Naval was right. Code, content, and capital are the only forms of leverage that compound without proportional effort. Here is what that means for agency owners in 2026.
Linear effort produces linear results. Automation compounds. Here is how 90 days of disciplined automation work creates a moat that five years of managed headcount cannot build.
If your business runs on platforms you don't control, you don't own the business. Here is what stack ownership means in 2026 and why it compounds like no other moat.
Book a free 30-minute call to scope your project. Fixed pricing, transparent timelines.
| Task category | AI replaces? | Specific knowledge required? |
|---|
| Market research and data aggregation | Yes, fully | No |
| Initial draft copywriting | Yes, substantially | Sometimes (voice, judgment) |
| SEO audit and technical checklist | Yes, substantially | Occasionally (priority judgment) |
| Strategic positioning for a client | Partially (data gathering) | Yes (judgment, context) |
| Client relationship management | No | Yes (trust, reading the room) |
| Crisis communication | No | Yes (judgment, consequences) |
| Identifying which clients to turn down | No | Yes (specific knowledge) |
| Building and maintaining reputation | No | Yes (skin in the game) |
