Working with AI: the new division of labor

Consider how a chef works with kitchen equipment. A food processor can chop vegetables in seconds. But the chef decides what to chop, how fine, and when.

• The machine handles the mechanical part.
• The human handles the judgment.

AI works the same way. It can generate text, analyze data, and produce variations far faster than any human. But it cannot tell if the output is good, appropriate, or true. That judgment still requires a human.

Generating variations

Ask AI for ten email subject lines, and it produces them in seconds. Ask for twenty, and the time barely changes. The generation cost of more options is nearly zero. The selection cost is not; scanning fifty options takes more time than scanning five. But skimming is faster than creating from scratch, so the trade-off usually works.

Why this matters:

• A copywriter who writes three headlines has fewer choices
• Someone who generates fifty and picks the best five has more to work with
• The skill shifts from writing to recognizing quality quickly

Processing at scale

Tasks that would take humans hours can happen in minutes:

• Reading through hundreds of documents
• Summarizing long reports
• Categorizing large datasets

This is not magic. AI is pattern matching at high speed. For tasks where pattern matching is the core work, that speed creates real value.

First drafts and starting points

Starting from a blank page is psychologically difficult. AI can produce a rough version quickly — not perfect, but something to react to.

Knowing if something is true

AI generates text by predicting what words should come next. It does not check whether those predictions are factually accurate.

What this means in practice:

• AI will confidently state things that are wrong
• It will invent statistics and create fake quotes
• The text sounds authoritative regardless of whether it is true

Key insight: Anything with specific facts, names, dates, numbers, needs verification.

Understanding context it was not given

AI knows nothing about your specific situation unless you tell it:

• Your company's culture
• Your customer's history
• The political dynamics in your organization

The more situation-specific the task, the more information AI needs.

Judgment that requires experience

Questions AI struggles with:

• Is this email too aggressive?
• Will this strategy offend customers?
• Is this joke appropriate for the audience?

These require understanding cultural nuance, relationship history, organizational norms. AI guesses based on general patterns, not your specific context.

The generation phase

Instead of creating one version, AI creates many:

• Ten subject lines
• Five opening paragraphs
• Twenty product names

This phase benefits from broad instructions. Give me options rather than give me the perfect answer. Perfection is not the goal, optionality is.

The selection phase

With options generated, humans evaluate:

• Which ones feel right?
• Which ones miss the mark?
• What elements from different options could combine?

Knowing what good looks like. Understanding what will resonate with the specific audience.

The refinement phase

Selected options get improved:

• Editing AI output
• Combining elements from multiple versions
• Using the AI draft as inspiration for something new

The verification phase

Before anything goes out:

• Facts get checked
• Claims get verified
• Appropriateness gets evaluated by someone who understands the context

This step is often skipped because AI output looks polished. But polish is not accuracy.

The generation-selection model fits many tasks but not all.

Tasks where verification costs more than creation

Some work requires such careful checking that validating AI output takes longer than doing it originally:

• Legal documents where every clause matters
• Technical writing where accuracy is critical
• Research where every claim needs a source

Key insight: If verification burden exceeds time saved in generation, the approach breaks down.

Tasks where effort itself matters

A handwritten thank-you note carries meaning because someone took the time. A personally crafted apology matters because the effort signals sincerity.

Using AI for these tasks might produce adequate text while destroying what made the gesture valuable.

Situations requiring genuine expertise

The selection phase requires knowing what good looks like. Without that knowledge, choosing among AI options becomes guessing.

Examples:

• Someone without legal training cannot pick the best contract language
• Someone without medical knowledge cannot evaluate health advice

Work where originality is the point

AI outputs cluster around patterns in its training data. It produces variations on what already exists.

For work that needs to be genuinely new:

• Breakthrough ideas
• Distinctive creative voices
• Contrarian perspectives

This tendency toward the average works against the goal. AI can contribute starting points, but differentiation must come from human contribution.