Concepts
AI Automation Work Technology

Codifiability Threshold

Origin : Frey & Osborne (2013) / Arntz-Gregory-Zierahn OECD (2016) — synthesis

What determines whether a task shifts to automation is not its difficulty, but its codifiability — can its rules, patterns, or sequences be extracted and taught to a machine? Four technological breakthroughs have successively raised this threshold.

Radiology seemed protected by years of training. Legal research, by legal culture. Writing, by creativity. All crossed the threshold — not because they were simple, but because they became codifiable. That threshold is what determines what shifts, not apparent complexity.

Principle

The codifiability threshold is the point at which a task can be learned and executed by an automated system. What determines the crossing is not the difficulty of the task, but the possibility of extracting its underlying rules, patterns, or sequences.

Key implication: a very complex task can cross the threshold quickly if it is structurally codifiable. A simple task can resist for a long time if it is irreducibly relational.

The 4 Types of Codifiability

1 — Deterministic Rules (1980-2000)

Tasks expressible as explicit rules: if X then Y.

Examples: data entry, calculation, sorting, batch processing.

Once you can write the algorithm, the task shifts. This is the industrial automation wave.

2 — Statistical Patterns (2010-2020)

Tasks not expressible as rules, but whose patterns are detectable in data.

Examples: fraud detection, image recognition, recommendations, structured-data diagnostics.

The model extracts patterns without anyone knowing how to formulate them explicitly. The task shifts without a single rule being written.

3 — Discursive Reasoning (2020-2023)

Tasks related to language, contextualization, reasoning over text.

Examples: writing, synthesis, code, analysis, customer support, translation.

LLMs showed that these apparently “human” tasks are codifiable through learning on massive volumes of human production.

4 — Complex Sequences (2024-2026)

Tasks protected not by the difficulty of an isolated step, but by the complexity of the sequence — sustained judgment over a chain of actions.

Examples: complete legal research, integrated radiology, multi-step workflows.

Autonomous agents have this continuity of execution. Protection through chain complexity collapses.

The Apparent Complexity Paradox

A task can seem very complex and still cross the threshold quickly. What matters is not perceived complexity, but the nature of what makes it difficult:

Practical Question

For each task in your role, identify its codifiability type:

  1. Is it expressible as rules? → Type 1, already shifted
  2. Does it have learnable patterns in data? → Type 2, already shifted
  3. Is it primarily discursive? → Type 3, shifted or in progress
  4. Is it a complex sequence? → Type 4, next wave
  5. Does it require irreducible human presence? → Structurally resists

Sources

Concepts