What AI Actually Does

Core Concepts

AI operates on tasks, not jobs

When people say "AI will take jobs," they are compressing a more precise mechanism. Research consistently shows that automation occurs at the task level, not the job level. An occupation is a bundle of heterogeneous tasks — some of which may be susceptible to automation, others resistant, and others better suited to augmentation. Treating a job as a single unit produces imprecise predictions; treating it as a collection of distinct tasks produces much more accurate ones.

A radiologist's job, for instance, includes image pattern recognition (high automation potential), clinical communication with patients (low automation potential), and treatment planning in collaboration with other clinicians (mixed). AI does not eliminate the radiologist; it reshapes which tasks take more or less of their time.

The routine/non-routine partition

The foundational framework for analyzing automation comes from labor economics. The task-based framework partitions occupational work along two axes:

• Routine vs. non-routine: Can the task be accomplished through a set of explicit, codifiable rules?
• Cognitive vs. manual: Does the task involve information processing and judgment, or physical action?

This creates four quadrants:

Fig 1

The practical prediction from this framework: computer capital substitutes for workers in routine cognitive and manual tasks (top-left and bottom-left quadrants), while it complements workers performing non-routine problem-solving and complex communication tasks (right side).

The historical shift: from skill-bias to task-content

It is worth understanding how this framework emerged, because it explains a real shift in automation's logic. During the 1980s and 1990s, computerization appeared to favor higher-educated workers — more college graduates were employed, and their wages rose relative to workers without degrees. This was called skill-biased technological change (SBTC).

From the mid-1990s onward, however, research documented a different pattern: computers were substituting for routine tasks — regardless of whether those tasks were performed by educated or uneducated workers. A data entry clerk with a college degree and a factory line worker without one both saw their routine tasks automated. The organizing principle had shifted from the worker's education level to the task's routineness. This is routine-biased technological change (RBTC).

This distinction is not just historical trivia — it is the foundation for understanding why the current AI wave feels different, which we will come to shortly.

Automation vs. augmentation: not the same thing

The same technology can interact with a task in two fundamentally different ways:

• Automation: Capital (the AI system) substitutes for human labor in executing the task. The human is removed from the loop.
• Augmentation: Capital and human labor work together. The AI amplifies what the human can do without replacing them.

Research across 1,500 workers in 104 occupations finds that workers themselves prefer higher levels of human agency than experts deemed technologically necessary on nearly half (47.5%) of tasks. The dominant worker-preferred mode in 47 out of 104 occupations was equal partnership — human-agent collaboration rather than full automation.

Automation is not solely a technological inevitability. It reflects managerial and organizational decisions about how to deploy technology.

This is not a minor point. It means that when you observe a given AI deployment in practice — a customer service chatbot, a coding assistant, an AI-assisted diagnostic tool — the question of whether it is automating or augmenting is partly a design and organizational choice, not just a technological one.

What tasks resist automation

Some task categories remain resistant to AI automation despite significant technological advances. Research identifies four main properties that confer resistance:

1. Physical and unstructured contexts — tasks requiring adaptation to novel physical environments.
2. Novel situational judgment — decisions that cannot be reduced to patterns in prior data.
3. Creative and generative work involving genuine originality — though this boundary is actively shifting.
4. Emotional and social intelligence — empathy, trust-building, nuanced interpersonal navigation.

Roles dominated by these task types — sociologists, management analysts, roles with high interpersonal components — tend to experience AI as an augmenting force rather than a substituting one.

The LLM exception: when the frontier moves

Here is where the established framework gets complicated. Large language models and generative AI represent a departure from prior automation waves in a specific way: they are targeting non-routine cognitive tasks — the top-right quadrant that was previously considered protected.

Earlier automation displaced manual and routine cognitive workers. Generative AI is demonstrating capability to perform or augment complex reasoning, creative generation, and judgment-requiring tasks in fields like law, journalism, and software development — industries where workers apply specialized training to non-routine problem solving.

This does not mean all non-routine cognitive work is now automatable. It means the boundary has moved, and the task-based framework's core assumption — that computational systems complement rather than substitute for abstract analytical work — no longer holds universally.

AI adoption is not uniform

One final conceptual anchor: AI adoption does not hit all occupations, firms, or geographies equally. Across 38 OECD countries, a one standard deviation increase in AI adoption correlates with a 2.3% reduction in employment in routine cognitive occupations — but a 1.8% increase in employment requiring complex problem solving and interpersonal skills.

Wage effects are asymmetric too: workers in the top income quintile experience wage gains, while middle quintile workers face modest declines. This heterogeneity matters because aggregate statistics ("AI will create X million jobs and destroy Y million") can obscure the distributional effects that matter most for specific people and sectors.

Researchers use occupational task databases — particularly the U.S. Department of Labor's O*NET — to classify and measure exposure to automation at the task level. This granular analysis predicts occupational AI exposure with greater precision than job-level or sector-level approaches.

What AI is doing to the occupational map

AI adoption is also creating new roles that did not previously exist. Employment in AI-related roles — software developers, data scientists, database architects — is growing significantly faster than average (17.9% projected growth for software developers between 2023 and 2033, versus 4.0% average). New occupational categories are emerging: AI trainers, prompt engineers, AI ethicists, explainability specialists.

The net balance of displacement and creation is a genuinely contested empirical question — and one that varies by timeframe, region, and which workers you are asking about. Later modules will examine the evidence on both sides.

---

Analogy Bridge

If you want a mental model for the automation/augmentation distinction, consider power tools.

A power saw can replace a handsaw entirely for cutting lumber (automation). But a surgeon's powered drill does not replace the surgeon — it amplifies precision in tasks the surgeon still controls (augmentation). The same class of technology, two different relationships with human labor.

Now extend the analogy: imagine a power tool that can handle not just physical cutting, but also planning the cut, adapting the design when the wood grain changes unexpectedly, and explaining its reasoning to the client. That is roughly what generative AI is doing to cognitive work — it is a power tool for thought that can increasingly handle parts of the task that previously required a skilled human throughout.

The analogy breaks down at the edges — AI is more adaptive and generative than any physical tool. But as a starting intuition for the automation/augmentation split, it is a reliable frame.

---

Compare & Contrast

Automation vs. augmentation at the task level

Dimension	Automation	Augmentation
Human role	Removed or minimized	Remains central; AI extends capacity
Locus of control	AI system	Human, with AI assistance
Worker preference (research finding)	Often lower than expert assumptions	Preferred mode in most occupations studied
Risk profile	Displacement if task constitutes bulk of role	Generally productivity-enhancing
Example	AI chatbot handles customer query end-to-end	AI draft + human editor for written content

Routine-biased vs. generative AI automation

Dimension	Routine-biased (RBTC era)	Generative AI era
Target tasks	Routine cognitive and manual	Non-routine cognitive, increasingly
Affected workers	Clerical, administrative, factory	Knowledge workers, professionals
Framework status	Well-validated, stable	Framework under revision
Complementarity assumption	Holds for non-routine cognitive	No longer universal
Example	Spreadsheet software replacing bookkeepers	LLM drafting legal briefs