What Does AI Actually Replace?

A centric figure relies on a fixed radius and a single pivot point

The geometry of a drawn figure is defined by its organizing principle, not by the instrument pressed against the paper. A circle is not defined by ink; it is defined by the fixed distance between a central anchor and a sweeping perimeter. To construct that figure, a draftsman requires three elements: an origin, a defined radius, and a tool capable of translating that radius into a stroke. The shape is the structural goal. The compass is merely the instrument that executes the curve.

When geometric drafting moved from manual scribing to mechanical tools, the underlying mathematics did not alter. The center point remained anchored where the draftsman placed it. The required radius remained dictated by the engineering specifications of the drawing. What changed was exclusively the friction of drawing the line. A mechanical compass eliminated hand tremors and reduced execution time, but it did not choose where to place the point, nor did it decide whether the architectural plan required a circle or a square.

Modern enterprise workflows operate on identical geometric principles. An enterprise process is an assembly of distinct operational movements organized around a strategic intent. Every professional role—whether in software engineering, legal practice, financial analysis, or customer operations—consists of an underlying architecture driven by specific functional sub-tasks. The execution of those sub-tasks is the mechanical sweep of the instrument. The structural design, the strategic boundary conditions, and the ownership of the output constitute the shape itself.

Misunderstanding technology always begins at the tool head. By confusing the instrument that sweeps the arc with the system that defines the radius, organizations mistake faster execution for structural transformation.

Viewing automation literally leads to the illusion of total role substitution

When an automated tool sweeps a line across a blueprint in milliseconds, observers routinely conclude that the draftsman has become obsolete. Read literally against artificial intelligence, a single software model capable of generating syntactically correct code or drafting a legal memo looks like an entity replacing a human professional wholesale.

This literal reading stems from observing surface outputs rather than structural mechanics. When a generative model writes a three-hundred-line block of routine code, the visible labor of writing disappears. Because traditional labor models equate spent hours with professional output, the elimination of those hours is misread as the elimination of the job.

In enterprise software engineering, developers routinely use automated tools to generate 30% to 40% of their daily code volume. A developer using GitHub Copilot completes an HTTP server implementation task 55.8% faster than an unassisted peer, according to empirical research by Peng et al. (2023). Observational observers look at a 55.8% speed increase and declare that half the engineering team is no longer required.

That conclusion misinterprets what actually occurred. The tool automated the mechanical translation of intent into syntax—the retrieval of API patterns, the writing of standard loops, and the setup of repetitive test skeletons. It did not formulate the system architecture, establish security boundaries, model domain data, or take responsibility for production stability.

The software did not redesign the building; it turned the compass faster. What looks like a single tool replacing a single job is actually a high-speed instrument executing a specific mechanical sub-construction inside a larger, unchanged architectural framework.

Artificial intelligence replaces specific constructions rather than complete roles

What AI actually replaces is not roles or jobs wholesale, but specific constructions. Not the circle, not the square—the compass. What AI replaces is the tool that draws the shape, not the shape itself.

A job is not a monolithic block of matter. It is a composite assembly of four distinct operational constructions: retrieval, synthesis, drafting, and judgment. Generative intelligence is a specialized engine for the first three. It substitutes for the mechanical friction of finding information, combining disparate inputs, and formatting them into standard structures. It does not replace the human anchor that defines why those constructions exist or whether their output is valid.

Critics frequently object that if an automated tool absorbs drafting, retrieval, and initial synthesis—which often account for 70% of a worker’s daily hours—the total demand for human labor must collapse proportionally. This objection relies on a static view of economic demand.

When the marginal cost of executing a construction drops toward zero, total demand for the underlying activity does not shrink; it expands. This is Jevons Paradox applied to knowledge work. Lowering the cost of drafting boilerplate code or scanning legal documents does not mean organizations buy fewer software architectures or execute fewer legal transactions. It means they build larger systems and analyze complex data rooms that were previously cost-prohibitive.

Furthermore, removing human oversight from the judgment construction yields immediate structural failures. When sub-tasks are automated without human anchoring, systemic quality collapses. What changes under automation is the point of human leverage: execution capacity ceases to be the operational constraint, shifting the entire bottleneck to the clarity of systemic intent and the quality of human judgment.

Enterprise tasks decompose into four discrete operational constructions

To understand where artificial intelligence operates cleanly and where it breaks down, enterprise operations must be evaluated case-by-case across their four core sub-constructions.

+-----------------------------------------------------------------------+
|                      THE FOUR CONSTRUCTIONS                           |
+-------------------+---------------------------------------------------+
| RETRIEVAL         | Searching databases, fetching files, parsing docs |
+-------------------+---------------------------------------------------+
| SYNTHESIS         | Aggregating sources, summarizing, formatting      |
+-------------------+---------------------------------------------------+
| DRAFTING          | Generating boilerplate, syntax, initial text      |
+-------------------+---------------------------------------------------+
| JUDGMENT          | Verifying logic, evaluating risk, accountability  |
+-------------------+---------------------------------------------------+

1. Retrieval & Synthesis

In high-volume customer support operations, the primary friction is searching documentation databases and synthesizing context-specific answers. A landmark study of 5,179 customer support agents by Brynjolfsson et al. (2023) demonstrated that generative AI assistants increased issue resolutions per hour by 14% on average. For novice workers, productivity surged by 34% to 35%.

Crucially, experienced workers saw zero to negligible speed gains. The tool did not replace expert diagnostic capacity; it replaced the retrieval and initial response synthesis constructions that inexperienced workers previously spent months learning.

2. Drafting and Formatting Mechanics

For professional writing tasks such as press releases, reports, and grants, an MIT study by Noy & Zhang (2023) showed that generative tools reduced task completion time by 40% while elevating output quality by 18%.

The software substituted directly for first-draft mechanics—the friction of turning structured ideas into polished prose. By taking over the drafting construction, the tool compressed the time spent moving the pen, allowing the professional to focus on refining the underlying thesis.

3. Pattern-Matching Scan

In medical diagnostics, routine low-risk scanning is a high-volume pattern-matching construction. A prospective trial of 31,301 screening mammograms published in Nature Medicine and analyzed by Elías-Cabot et al. (2026) found that autonomous AI triage of low-risk scans reduced radiologist reading workload by 63.6% while increasing cancer detection rates by 15.2%.

The AI replaced the basic exclusion scan—drawing the routine line—reserving human radiologists for complex, ambiguous diagnostic judgments.

4. Qualitative Judgment and Logic

The boundary of automation becomes sharpest when execution hits qualitative judgment. A field experiment conducted with 758 Boston Consulting Group consultants by Dell’Acqua et al. (2023) mapped this capability boundary, termed the “jagged technological frontier.”

For tasks inside the frontier (creative drafting and analytical synthesis), consultants using AI completed 12.2% more tasks, 25.1% faster, with 40% higher quality. But for tasks outside the frontier—requiring nuanced logic and qualitative business judgment—consultants using AI performed 19 percentage points worse than unassisted colleagues.

BCG CONSULTANT PERFORMANCE (Dell'Acqua et al., 2023)
Inside Frontier (Drafting/Synthesis):  [+40% Quality Boost      ]
Outside Frontier (Judgment/Logic):      [-19 Percentage Points    ]

When professionals treated the tool as a replacement for judgment rather than a tool for drafting, performance crashed. The compass cannot decide where the center point belongs.

Erasing the draftsman because the instrument turned faster exposes the enterprise to structural collapse

Assuming that an entire operational function is replaced simply because its mechanical constructions have been automated leads directly to institutional failure. Enterprise leadership teams routinely make this mistake by measuring tool speed and concluding they can discard the operational anchor.

Consider the customer support rollout at fintech firm Klarna. In February 2024, the company announced that its AI assistant handled 2.3 million customer chats in its first month—67% of total support volume—doing the equivalent work of 700 full-time agents and reducing average resolution times from 11 minutes to under 2 minutes, as detailed in the Klarna Global Press Release (2024) and an OpenAI Case Study.

The tool successfully automated specific constructions: order status retrieval, canned policy synthesis, and basic account routing. However, operating on the belief that these metrics meant the entire customer service function had been solved, headcount was reduced aggressively.

By mid-2025, the limits of replacing human oversight became clear. As analyzed by the AI Insight Lab Analysis, Klarna publicly adjusted its trajectory and reopened hiring for human support agents. While automated text generation excelled at routine queries, it failed at complex escalations, high-empathy customer disputes, and ambiguous fraud cases. The company had mistaken the rapid movement of the compass for the completion of the entire service shape.

A contrasting approach appears in professional legal services. Magic Circle law firm Allen & Overy deployed Harvey, a domain-specific legal model, to over 3,500 attorneys across 43 offices, as documented in the Allen & Overy (2023) announcement.

Lawyers processed tens of thousands of queries to scan 500-page data rooms, identify indemnification clauses, and draft regulatory filings. The firm did not eliminate its legal associates. It recognized that Harvey substituted for sub-constructions of document review and boilerplate drafting. Strategic risk calculation, client negotiation, and legal liability remained anchored entirely to human partners. The firm accelerated its compass without discarding its engineers.

Industry analysts counter that advanced reasoning models and multi-step autonomous agents are moving beyond simple drafting toward long-horizon planning, effectively choosing the shape themselves. This argument mistakes extended execution chains for strategic intent.

An autonomous agent can execute a twenty-step software deployment or construct a multi-tiered legal contract without human intervention. In doing so, it operates a more complex, self-correcting compass. But it does not possess legal standing, it cannot bear fiduciary responsibility, and it cannot answer for a failed compliance audit in a court of law. Intent, risk ownership, and operational accountability cannot be offloaded to an algorithmic model.

When an organization replaces the draftsman because the instrument learned to sweep the arc automatically, the drawings come out faster, but the buildings collapse under their own weight.

The venture studio used to be a room full of people; the next one is an operating system.