Vol. III · Essays on work after AI
Saturday, October 3, 2026
evolving.work
Judgment · Essay № 013

Four Skills for the AI Era

The capabilities that matter most in an AI-saturated job are not technical.

H Héctor Reinoso Apr 5, 2026 12 minute read Share · Save

You Cannot Judge a Wine You Have Never Tasted

Pour a $15 bottle and a $150 bottle into identical glasses, hand them to most people, and they will guess wrong. Not for lack of access to wine. For lack of a trained palate.

AI output works the same way. It arrives fluent, structured, confident. To anyone without the judgment to tell the difference, the brilliant draft and the plausible-but-wrong one look identical. I see it most weeks: someone shows me an AI draft, visibly pleased, the prose immaculate. Three claims in, something is confidently, expensively false.

The story you have been told is that AI democratizes expertise — point it at a problem and it performs like an expert, no apprenticeship required. The data tells a more uncomfortable story. When Harvard researchers gave 640 small-business entrepreneurs in Kenya an AI advisor and tracked real revenue over months, the AI group showed no overall advantage over those handed a written guide. The average hid a split: high performers improved 10-15%, while low performers declined 8% — actively worse for having used it [3].

The tool meant to lift everyone made the strong stronger and the weak poorer. That gap points somewhere specific. When execution stops being scarce, judgment becomes the scarce thing [5]. That is the verification economy in one line: AI output is noise until someone with real domain knowledge confirms it holds. The only open question is what that someone has to know how to do.

It collapses into four skills. None of them is technical. Every one runs on the domain expertise AI is supposedly making obsolete — which is either a cruel joke or the most reassuring news of the decade, depending on whether you have any.

Skill One: See the Right Problem

AI will answer any question you ask it, confidently, including the wrong one. It generates at scale but cannot decide which questions matter. That decision sits upstream of everything else, and it determines whether the work that follows creates value or destroys it.

The Kenya entrepreneurs make the mechanism visible. Low performers asked generic questions and took the confident, generic advice at face value. High performers spent their first move on framing — naming the real problem in terms of their specific market before asking AI for anything [3]. The danger is not visible failure. It is misaligned success: an eloquent answer to a problem that does not exist, looking like progress the whole way down. Failure announces itself. Misaligned success compounds in silence.

“Problem identification” sounds too soft to teach, which is exactly why it falls off the syllabus. But it is concrete. State the outcome you want to change without naming the tool. Describe what a wrong answer would look like. Ask why, five layers down, before you write a single prompt. The failure mode is solution-first thinking — how can AI help me with X — long before anyone has checked that X is the problem at all.

The market has already priced this in. The World Economic Forum surveyed more than 1,000 employers across 55 economies and found analytical thinking ranked the number-one core skill, essential to seven in ten companies [4]. Employers are not hiring for prompt fluency. They are hiring for the judgment to see what is worth solving.

The practical version is a ratio. Spend a third of your effort defining the problem and the rest executing. Minutes of framing followed by hours of doing can leave you with polished work that answers the wrong question.

Skill Two: Say What You Mean

Specification is not typing better prompts. Prompt engineering asks what words to type; specification asks what AI needs to know about your domain, your constraints, and your standard of good. The first is a trick. The second is expertise.

Vague in, vague out. The default failure is under-specification — treating AI as a colleague who already shares your context, when it has none of it. The test is blunt: could a knowledgeable colleague execute your instructions without a single follow-up question? If not, neither can the model.

I have watched two people run the same task on the same model on the same morning and finish ten times apart. One types a line and spends the next hour nudging the output toward something usable. The other writes an eleven-minute brief, makes coffee, and comes back to finished work — then does five more before lunch. Same tool. The entire difference was in what they handed it.

This is why specification belongs to the domain expert, not the tool owner. Firms capture more value from AI when that expertise is distributed across the people who know the constraints that matter, rather than concentrated in IT — and investors reward the combination [8]. The person who can specify is the person who knows what good looks like, which is rarely the person who owns the software license. There is a quiet bonus, too: a tight spec is the same discipline that sharpens the memos you already write for humans. Learning to brief a machine makes you better at briefing people.

You might object that these four capabilities are too different to be one skill set — that specification is closer to requirements engineering, while tasting is connoisseurship. Fair. They develop along different paths. What unifies them is the root: each is an act of judgment drawing on the same domain expertise. A sommelier specifies a wine to a supplier and tastes the bottle that arrives using one body of knowledge. Different acts, one foundation.

Skill Three: Know Where the Frontier Breaks

In 2023, Harvard Business School and BCG ran one of the largest controlled experiments on AI in real knowledge work. They gave 758 consultants — analysts to partners — eighteen realistic tasks and randomly assigned them AI access, AI with a prompting primer, or nothing.

Inside AI’s range, the results read like the case for adoption: 12.2% more tasks completed, 25.1% faster, at 40% higher quality. Below-average consultants jumped 43% [1].

Then the tasks crossed an invisible line. On work outside AI’s capability, consultants using it were 19 percentage points less likely to produce correct answers than colleagues working without it [1]. The model handed them confident, well-structured, wrong solutions — and they trusted it, because nothing about a wrong answer announces itself as wrong.

The researchers named this the jagged technological frontier. AI is uneven across tasks of similar apparent difficulty: brilliant at ideation and market analysis, unreliable at judgment that requires reconciling conflicting data. From the outside, the two jobs look identical. The consultants who got hurt were not careless; they treated the tool as uniformly capable and never felt the edge as they stepped over it. The differentiator was not AI access or prompting skill. It was knowing where the edge sat.

The same pattern surfaces wherever the work hinges on judgment rather than prediction. In structured university debates, high-ability debaters improved roughly 12% with AI while low-ability ones barely moved — because winning a debate turns on weighing credibility and rebuttal, not forecasting [10]. Where judgment dominates, AI widens the gap instead of closing it.

So map your own frontier into three zones. Delegate: tasks where AI reliably clears your bar. Collaborate: useful starting points that still need real editing. Retain: work where AI output is unreliable or dangerous to trust. Then keep redrawing the map — this is the first workforce skill that expires quarterly. The frontier moves every time a model ships, so the map is a standing practice, not a certificate you frame once and hang on the wall.

The moving frontier also changes how much judgment the work requires. They imagine a rising tide: AI improves, the patch of human work shrinks, the water eventually covers everything. But every task the model absorbs creates a new boundary, and every boundary is a fresh judgment call about whether to trust what comes back across it. Think of AI capability as a bubble — the work that matters happens at its surface. As the bubble expands, that surface gets larger, not smaller. More calls to make, not fewer.

Skill Four: The Taster’s Advantage

In August 2025, three separate federal courts sanctioned attorneys in a single month. The offense was identical: legal briefs citing cases that did not exist.

The drafting had been well inside the frontier. The prose was clean, the reasoning plausible, the citations formatted to perfection. They were also invented. The lawyers had outsourced not just the writing but the verification — and crossed the frontier without noticing [7]. This was not three rogue practitioners having a bad month. Stanford measured legal-AI hallucination at 17-43% for specialized tools and 58-88% for general models, and found 83% of legal professionals report meeting fabricated case law [7].

Tasting is the skill that catches this. It is judgment trained through practice, comparison, and deep exposure — exactly how a palate is built [9]. A sommelier does not send each glass to a lab; she has calibrated against thousands until she can taste what is off in a second. A novice lawyer cannot tell a real citation from a hallucinated one for the same reason a novice cannot tell the $15 bottle from the $150. You cannot judge a wine you have never tasted.

You may suspect this advantage is temporary — that AI is developing its own judgment and will close the gap soon. Benchmarks do climb fast, so take the strongest version of the worry. On Humanity’s Last Exam — 2,500 expert-built questions across more than 100 domains — the best model scored 37.5% against human experts near 90%, a 53-point gap [6]. And MIT research found that halving an error rate demands orders of magnitude more compute [11]. The last mile of quality is economically irrational to automate. Human tasting appreciates; it does not depreciate.

The discipline is contrarian by design. When was the last time you said no to AI? Your rejections are worth more than your prompts — they are where domain knowledge actually shows up.

So It Only Helps People Who Are Already Experts?

One objection should keep me honest, and it deserves the full version: this framework privileges existing experts and bolts the door behind them. AI was supposed to let newcomers skip the apprenticeship — and the data backs that promise. Inside the frontier, below-average BCG consultants gained 43% [1], far more than the top performers. In call centers, where the work is mostly prediction, novices improved 34% and the gap between weak and strong workers narrowed [2]. AI genuinely democratizes — within the frontier.

So the line has to be redrawn, not erased. Beyond the frontier, where ambiguity and stakes live, the Kenya study shows AI without judgment destroys value rather than spreading it [3]. Telling a newcomer they can skip domain expertise is the real disservice — they will hit the frontier at full speed and never feel it.

But taste is learnable, not innate [9]. The four skills are a development roadmap, not a velvet rope. Used well, AI becomes an accelerated apprenticeship — a faster path to the expertise that earns judgment, not a substitute for earning it. Two years of AI-native work, spent crossing the frontier on purpose, may build sharper pattern recognition than eight years of the traditional climb. The door is not bolted; the stairs just got steeper and faster at once.

What to Do With This

The four skills — see the right problem, say what you mean, know where the frontier breaks, taste the result — are all judgment skills. Every one of them runs on the domain expertise everyone assumes AI is replacing. That irony is not a coincidence. It is the entire point.

One caveat: for low-stakes work well inside the frontier, skip the ceremony and use the tool. Not every email needs an eleven-minute spec.

For everything that matters, the move is concrete. If you lead L&D, audit your teams against these four, not against tool proficiency. If you hire, write the job description around judgment and stop screening for prompt tricks. If you work alone, it starts this week: pick the skill you are weakest at and run your next real task through it deliberately — name a problem without naming the tool, or find the one thing AI got subtly wrong and say exactly why.

The AI is table stakes. Anyone can rent the same model you can. What they cannot rent is your palate — and palates, unlike model subscriptions, only grow more valuable as the wine gets better.

References

organizationsAI transformation
About the author
Héctor Reinoso writes about how organizations should be designed when intelligence becomes abundant. He works on enterprise AI adoption inside a large industrial organization, after 25 years as a strategy consultant and transformation executive across four continents. Harvard MBA, industrial engineer, and still writes the code he writes about. More →
Continue reading
Strategy · Oct 3, 2026

Twelve Agents, €200, and a Consultant With a Name

When agents make research cheaper, consulting must earn its fee through the structure of the analysis, the quality of judgment, and clear accountability.
Architecture · Oct 1, 2026

From Writing Code to Managing Agents

Coding agents let developers work with larger units than individual lines. The skill is knowing how to delegate, check, and investigate.
Org Design · Sep 24, 2026

When Time Collapses, What Breaks First?

Time was an invisible governance system in organizations, not just a constraint.
Trust · Sep 7, 2026

The 17% That Matters

AI reasoning and output operate as semi-independent processes: a model can identify the correct answer in its reasoning and then produce a recommendation that contradicts it.
Knowledge · Aug 19, 2026

Is YouTube Still Learning — Or Just Repetition?

When content production cost collapses, volume rises but novelty compresses.
Strategy · Jul 31, 2026

When Intelligence Becomes Cheap, What Becomes Expensive?

AI is driving the marginal cost of intelligence toward zero, but abundance does not eliminate scarcity — it migrates it.