Is the 10,000-hour rule absolute?

No. It's a heuristic popularized by Malcolm Gladwell based on research by K. Anders Ericsson. It doesn't mean that below 10,000 hours a person has no skill — it means 'world-class expertise' requires deliberate practice time. Applied to Generative AI in 2026, it shows that no one can yet have full expertise.

So should I never hire an 'AI expert'?

You should. But with calibrated humility. In 2026, 'Generative AI expert' means 'someone who has been studying hard for 2-3 years and is still learning'. Anyone who sells themselves as having full expertise is signaling either ignorance or bad faith. Look for those who admit they're still learning — those are the best ones.

How do I evaluate who actually knows their stuff?

Four tests: (1) do they personally build their own agents and tools? (2) can they explain architecture conceptually without excessive jargon? (3) do they admit what they don't know? (4) do they have demonstrable hands-on work (GitHub, projects, technical posts), not just talk? Those who pass all 4 are the best available in 2026.

What about big consulting firms — Deloitte, McKinsey, Accenture?

They have excellent people, and they have people recycling 2018 slides. The problem is enormous heterogeneity within the same firm. Don't buy the brand — buy the specific person with a verifiable track record of real implementation.

How does Fhinck evaluate AI talent when hiring?

Through real hands-on work. Candidates must demonstrate a personal agent, prompts/MCPs they created, projects on GitHub. Those who only talk theory don't make it. Those who show real work, even as beginners, do.

Before Hiring a 'Generative AI Expert', Run This Simple Calculation

In 2026, the market is flooded with "Generative AI experts." Before signing the contract (or the consulting check), run an Ericsson calculation. The result is mathematically uncomfortable.

"Before hiring the next 'Generative AI expert,' run a simple calculation: Generative AI became accessible 39 months ago. Expertise requires at least 10,000 hours of deliberate practice. Even in the most theoretical perfect scenario, nobody in the world has logged more than 6,864 hours."

— Paulo Castello, February 2026

The Math Nobody Runs

Let's run it together:

ChatGPT launched in November 2022. The commercially useful generation of Generative AI (LLMs in production, usable prompts at scale, first agents) started there. Before that, Generative AI was academic research and niche technical tools — not something the average professional could learn at scale.

From November 2022 to February 2026 = 39 months.

The most optimistic theoretical scenario possible: someone studied Generative AI 8 hours a day, every single day, with no vacations, no weekends.

Calculation:

39 months × 30 days × 8 hours = 9,360 maximum theoretical hours

A slightly more realistic scenario (but still optimistic): 8h/day on business days, 22 days/month:

39 × 22 × 8 = 6,864 hours

The 10,000-hour rule (Anders Ericsson, popularized by Malcolm Gladwell) says that world-class expertise requires approximately 10,000 hours of deliberate practice.

Nobody in the world, mathematically, has logged more than 6,864 hours in Generative AI.

In practice, the real number is much lower:

Most people didn't study 8h/day
A significant portion of that time is passive reading (not deliberate practice)
Technology changed so much that part of 2023 knowledge is already obsolete in 2026

Realistic estimate for the world's top actual expert in Generative AI today: 4,000-5,000 hours of deliberate practice.

Nobody is a "fully formed expert" yet.

What This Means for C-Level Executives

Three practical implications for those evaluating talent or hiring AI consulting:

1. "Senior expert" in Generative AI doesn't exist — yet

Those who sell themselves as "senior" or "I have 10 years of experience in Generative AI" are lying (or misinformed). The commercially useful technology is 3 years old.

They may have 10 years in traditional AI (Machine Learning, Computer Vision, classic NLP). That's not the same thing. Concepts transfer partially. But the modern stack (LLMs, agents, MCP, orchestrators) was born post-2022.

2. The best ones admit they're still learning

Remember Karpathy's viral post from December 2025? "I've never felt so behind as a programmer."

If Karpathy — absolutely top 10 in the world — admits it, any "expert" who sells themselves as having full expertise is either ignorant or acting in bad faith.

Look for those who admit what they don't know. Those are usually the best.

3. Real hands-on work > impressive certifications

The relevant differentiation in 2026:

Those who talk about AI (consultants who teach courses, keynote speakers, article authors)
vs those who do AI (built their own agent, maintain open-source code, integrate MCP, have a real implemented case)

The two groups are not equivalent. For your AI First strategy, only the second one matters.

The 4 Tests for Evaluating AI Talent Today

When evaluating a candidate, consultant, or partner, apply:

Test 1 — Do they personally create?

Do they have a personal agent they built?
Do they maintain reusable prompts/contexts/MCPs?
Can they demonstrate live during the interview what they built?

Those who don't create, don't understand deeply.

Test 2 — Can they explain without excessive jargon?

Ask: "explain what NemoClaw changes in one sentence." Or: "difference between RAG and fine-tuning, in terms of when to use each."

Those who understand, explain simply. Those who repeat jargon, are recycling slides.

Test 3 — Do they admit what they don't know?

Ask: "what part of Generative AI do you still struggle with?" Those who answer "none" are signaling either ignorance or bad faith. Those who answer with technical humility, are trustworthy.

Test 4 — Do they have demonstrable hands-on work?

This could be:

Public GitHub with projects
Detailed technical posts (not just market opinion)
Presentations at technical meetups
Contributions to technical communities (Anthropic Discord, OpenAI forum, Brazilian AI communities)

Without any of this, it's just talk. In 2026, talk is expensive.

What About Large Consulting Firms?

Deloitte, McKinsey, Accenture, BCG, EY — all have AI divisions. They have excellent people. And they have people recycling 2018 decks.

The problem is enormous heterogeneity within the same firm.

Practical rule: don't buy the brand. Buy the specific person.

Who will actually work on your project?
Does that person have real hands-on experience or will they just present slides?
What similar projects have they personally led?

If the answer is vague, you're paying Tier 1 prices and getting Tier 3 delivery.

How Fhinck Hires

For Fhinckers joining an AI First team:

Mandatory personal agent demonstration during the interview
Show prompts/MCPs created, with technical explanation of why
Real hands-on tested (not just conversation)
Those who only talk theory don't make it

Those who show real work, even as beginners who are studying hard, make it. Attitude > certifications.

Conclusion

In 2026, nobody is a "fully formed expert" in Generative AI. The math proves it.

The best ones admit they're learning. The worst ones sell themselves as complete authorities.

Companies that understand this nuance hire better, pay for better consulting, and don't fall into the trap of an "expert" who charges Tier 1 and delivers Tier 3.

Fhinck went through the AI First transition making this rigorous evaluation at every hire. Schedule a conversation if you want to understand how to apply this in your team.