Knowledge Work vs Professional Work
Relevant source files
The following files were used as context for generating this wiki page:
Knowledge work is defined as a superset where knowledge acts as both the raw material and the product, characterized by non-routine judgment. Professional work is a strict subset of knowledge work, gated by formal credentialing, exclusive jurisdiction, sign-off authority, and personal liability. In the context of AI agent evaluation, a central diagnostic issue is that AI capability claims are often stated in professional-work language but supported by knowledge-work benchmarks that lack the institutional layers of credentialing and accountability.
Sources: README.md:38-44
Defining the Taxonomy
The project utilizes an empirically derived 18-operation taxonomy of knowledge-work operations. These operations represent the smallest activity patterns that recur across various occupations and carry a recognizable professional construct. The distinction between general "knowledge work" and specific "professional work" is critical for establishing "handoff" validity—evaluating if an AI output can actually be acted upon within a real workflow involving sign-off and liability.
Sources: README.md:21-27, README.md:38-44
Data Flow for Task Classification
The system filters raw O*NET and ESCO data to isolate knowledge-work tasks from physical or routine clerical work through multi-stage screening.
The diagram shows the pipeline from raw occupational data to the categorized 18 operations, highlighting the additional "gate" required for professional work status.
Sources: README.md:104-122, scripts/round2_knowledge_screen.py:33-40
Triage and Screening Logic
The project employs a "Round-2 knowledge-work screen" to ensure tasks are primarily cognitive and produce judgment-bearing artifacts. This screen explicitly excludes tasks that are physical, manual, or routine clerical, even if they occur within a knowledge-work occupation.
Inclusion and Exclusion Criteria
| Category | Inclusion (Knowledge Work) | Exclusion (Physical/Routine) |
|---|---|---|
| Primary Focus | Cognitive, judgment-bearing | Direct physical, manual, or clinical |
| Output Type | Report, plan, analysis, advisory | Physical installation, repair, cleaning |
| Examples | Analyzing market conditions, developing models | Fitting orthopedic braces, repairing equipment |
| Administrative | Cognitive supervision, planning | Routine filing, answering phones |
Sources: scripts/round2_knowledge_screen.py:53-110, data/derived/work_pattern_labels_frozen.json:11-20
Implementation of the Screen
The round2_knowledge_screen.py script uses a Large Language Model (LLM) to categorize tasks.
# From scripts/round2_knowledge_screen.py:65-75
SYSTEM_PROMPT = """
Exclude tasks that are NOT primarily cognitive knowledge work...
EXCLUDE if the task is:
- a DIRECT PHYSICAL, MANUAL, CLINICAL, SURGICAL, OR LAB PROCEDURE
- PHYSICAL INSTALLATION, WIRING, PLACING, MOUNTING...
- ROUTINE CLERICAL or procurement EXECUTION
INCLUDE if the task is primarily cognitive, judgement-bearing, and produces an artifact or a recommendation.
"""
Sources: scripts/round2_knowledge_screen.py:53-110
Professional Work Operations
Professional work operations are those that carry a high "minimum task contract," requiring the worker to produce formal guidance or decisions that can be judged against explicit standards.
Core Knowledge-Work Operations
The following table highlights key operations identified in the project's taxonomy.
| Operation | Input | Output | Distinct From |
|---|---|---|---|
| Analysis | Raw evidence/data | Interpretive findings/memo | Appraisal (which uses rubrics) |
| Design | Requirements/constraints | Specifications/blueprints | Fabrication (which builds it) |
| Inspection | Subject + standards | Compliance verdict | Investigation (which is retrospective) |
| Advising | Specific client situation | Personalized recommendation | Teaching (which is group-based) |
Sources: data/derived/manual_labeling/operation_taxonomy.json:20-150
The Validity Gap
Evaluation of AI agents often fails at the "professional work" boundary due to the "handoff" validity gap.
The sequence diagram illustrates that while an AI can perform knowledge work (generating the artifact), the transition to professional work requires a human professional to assume liability and provide authorized sign-off.
Sources: README.md:29-36, README.md:145-155
Scoping ESCO and O*NET
The project uses specific filters to define the scope of knowledge work within international standards.
- ISCO Major Groups 1-3: Managers, Professionals, and Technicians/Associate Professionals define the knowledge-work scope.
- ESCO Skill Types: Only "skill/competence" items linked as essential to these ISCO groups are included, excluding "knowledge" items which describe subject-matter domains rather than operations.
- O*NET Job Zones 3-5: These zones typically encompass occupations requiring medium to extensive preparation, aligning with knowledge and professional work.
Sources: scripts/esco/prepare_esco_items.py:15-30, README.md:130-140
Conclusion
The distinction between knowledge work and professional work is fundamental to the project's critique of AI agent evaluation. While AI agents demonstrate high capability in performing the cognitive operations of knowledge work (e.g., analysis, drafting, recording), current benchmarks fail to address the institutional and legal requirements of professional work. For an AI to be considered capable of professional-grade tasks, evaluation must move beyond "primitive" scores to include organizational validity factors like accountability, authorized use, and sign-off readiness.
Sources: README.md:21-44