Toward a Discipline of Knowledge Engineering

Identifying, extracting, and capturing knowledge from practitioners, systems, and organizational memory. This phase confronts the tacit knowledge problem directly: the most valuable organizational knowledge resists the extraction methods most commonly employed. Feigenbaum's (1977) knowledge acquisition research in expert systems identified this bottleneck four decades ago, and it remains the discipline's hardest problem. AI-assisted transcription, code analysis, and conversational extraction reduce the cost of the tacit-to-explicit conversion, but the validation question (does the extracted knowledge accurately represent what the practitioner knows?) requires disciplinary methods that do not yet have an institutional home.

Structuring captured knowledge in systematic representations (taxonomies, ontologies, metadata schemas, content models) that make it machine-processable and human-discoverable. This phase applies the classification expertise that Bowker and Star (1999) analyzed: every classification decision embeds assumptions about who will use the knowledge and how, and these assumptions persist as invisible infrastructure long after the decision is forgotten. The knowledge architecture framework's design primitives (structural decomposability, semantic self-description, temporal binding, modality independence) apply with particular force here, and the heterogeneous-reasoner problem makes formalization decisions consequential across consumer classes that the formalizer cannot anticipate.

Arranging formalized knowledge in structures that support discovery, navigation, and consumption across different consumer classes. Procida's Diataxis framework provides the strongest existing methodology: different cognitive needs (learning, problem-solving, reference consultation, conceptual understanding) require structurally different knowledge artifacts, and conflating them degrades effectiveness for every consumer. Organization encompasses the information architecture decisions that determine findability and the metadata strategies that determine whether knowledge is surfaced contextually or retrieved only through explicit search.

Delivering knowledge to its consumers in forms they can process and act upon. Transfer effectiveness depends on the match between the artifact's structure and the consumer's cognitive strategy (Gigerenzer 2000). Sweller's (1988) cognitive load theory provides the mechanism: unnecessary processing effort imposed by the artifact's format degrades the consumer's engagement with the knowledge itself. The knowledge architecture framework's extension of progressive disclosure (each resolution level must be complete and correct at its own depth) addresses a specific transfer failure observed across the research program. AI-mediated transfer (chatbots, retrieval-augmented generation, agent-facing tool schemas) introduces new transfer modalities that require the same deliberate design attention.

Keeping knowledge current, retiring obsolete knowledge, managing the lifecycle of knowledge artifacts as the systems they describe evolve. This phase is where the temporal unbinding failure documented in the legibility thesis originates: knowledge artifacts that describe a previous system state without signaling the divergence cause every downstream consumer to form beliefs based on stale evidence. Williamson's (1985) transaction cost analysis applies: the cost of maintaining knowledge infrastructure determines whether the infrastructure remains trustworthy over time, and under-investment in maintenance produces knowledge decay that compounds across every system that depends on the decaying artifact. AI can assist with staleness detection and consistency checking, but the governance decisions (what to update, what to retire, what to restructure) remain engineering judgments that require disciplinary expertise.

The knowledge architecture framework documented the heterogeneous-reasoner problem: knowledge artifacts designed for human consumers fail when autonomous agents arrive, because agents lack the contextual inference, ambiguity tolerance, and compensating heuristics that humans bring to poorly structured information. This is not a future concern. Agent systems consuming MCP tool schemas, API specifications, and structured documentation are encountering knowledge infrastructure designed for human developers, and the failure modes (misinterpreted constraints, suboptimal tool selection, cascading integration errors) are already measurable. Simultaneously, AI systems are producing knowledge artifacts, generated documentation, code summaries, meeting transcriptions, knowledge base entries, at a scale that exceeds the validation capacity of the humans who are nominally responsible for knowledge quality. The result is a knowledge layer that is growing faster than it can be curated, serving consumers whose requirements it was not designed to meet. Gigerenzer's (2000) ecological rationality framework explains the structural mechanism: cognitive strategies perform well only when matched to the informational structure of their environment, and knowledge infrastructure designed for one cognitive strategy degrades every other consumer's strategy. The design decisions being made now, about how AI-generated knowledge is validated, how agent-facing schemas are structured, how human and machine knowledge artifacts coexist, will constitute the knowledge substrate for the next generation of AI-augmented systems.

The AI transformation is not unique to knowledge work. SRE practitioners are integrating AI into monitoring, anomaly detection, and incident response. DevOps pipelines are incorporating AI-assisted code review, test generation, and deployment optimization. DevSecOps is adapting to AI-generated code that introduces novel vulnerability patterns. Software engineering itself is being reconceived as AI-augmented development. Each of these disciplines is absorbing the transformation from a position of existing formalization: they have established identities, career paths, standards, metrics, and institutional structures that provide the scaffolding for adaptation. Knowledge work has none of this scaffolding. The AI transformation is arriving at a function that is distributed across seven role titles with no shared identity, no standard methods, no quality metrics, and no career infrastructure that would allow practitioners to develop the expertise required to navigate the transition. The absence of formalization means that the discipline's response to AI will be fragmented, ad-hoc, and organizationally invisible, precisely when the stakes of the knowledge layer's design quality are highest. Nelson and Winter's (1982) analysis of organizational routines is instructive: an organization's adaptive capacity depends on the routines that encode its capabilities, and knowledge infrastructure is the medium through which routines become explicit, transferable, and improvable. A discipline that cannot articulate its own methods cannot adapt them.

Brooks's (1975) observation that coordination costs in knowledge-intensive work grow superlinearly with team size applies directly. As organizations grow, the number of systems, processes, and decisions requiring knowledge capture grows faster than the knowledge engineering capacity the organization employs. AI accelerates both sides of this equation: it increases the number of knowledge-producing and knowledge-consuming systems while simultaneously reducing the marginal cost of knowledge production in ways that mask the growing maintenance and validation burden. The gap between knowledge complexity and knowledge investment widens, producing the organizational phenomenon practitioners describe as 'documentation debt,' structurally analogous to technical debt but lacking the engineering framing that would make it visible as an investment problem. Ostrom's (1990) commons governance analysis provides the institutional frame: organizational knowledge is a common-pool resource subject to classic commons failure modes (depletion through neglect, degradation through unvalidated contribution, free-riding by those who consume without contributing). A formalized discipline provides the institutional infrastructure, the monitoring, standards, career incentives, and resource allocation mechanisms, that commons governance requires. Without this infrastructure, knowledge remains a commons without governance.

The claim that knowledge work can be practiced as engineering is not speculative. Procida's Diataxis framework demonstrates that documentation types correspond to distinct cognitive functions and can be categorized systematically. The Docs-as-Code movement demonstrates that knowledge artifacts can be subject to the same engineering practices (version control, peer review, automated testing, continuous deployment) as software artifacts. Splunk's product-first model (Gales et al. 2017) demonstrates that knowledge work integrates into product development as a first-class function. Tom Johnson's sustained work on API documentation methodology demonstrates the engineering depth that knowledge transfer requires at scale. These frameworks constitute existence proofs. What they lack is a unifying disciplinary identity that would allow organizations to invest in them as components of a single professional field rather than as independent initiatives that each require separate justification. The Write the Docs community provides the gathering point; the EKAW conference series provides academic engagement; the Diataxis framework provides systematic methodology. The pieces of a discipline are assembling. What they require is the formalization that converts a collection of practices into a profession with the institutional infrastructure to develop, transmit, and advance its own knowledge.

The research program documented how knowledge artifacts govern consumer behavior through their design choices: what descriptions include, what schemas omit, how errors are categorized, how version boundaries are signaled. When no discipline claims responsibility for these choices, they are made incidentally. Formalization makes the choices deliberate. A knowledge engineer designing an API reference considers whether its structure serves the multiple consumer classes that depend on it: human developers, autonomous agents, integration validators, AI-assisted development environments, and reasoner forms that have not yet emerged. The design primitives from the knowledge architecture framework (structural decomposability, semantic self-description, verification surface, temporal binding, modality independence, progressive disclosure, composability, adversarial resilience) become the engineering standards that knowledge artifacts are designed and evaluated against. The AI-specific extension is that knowledge artifacts must now be designed for simultaneous human and machine consumption, with explicit constraint metadata, structured verification surfaces, and temporal binding that machines can parse without the contextual inference that humans supply.

Polanyi's tacit knowledge problem is currently treated as an obstacle to documentation rather than as the central engineering challenge of a discipline. Technical writers work around it, knowledge managers acknowledge it, and organizations accept knowledge loss when practitioners leave as an unavoidable cost. AI changes the economics of externalization (the tacit-to-explicit conversion) without changing its difficulty. A formalized discipline treats the conversion problem as its core research agenda, developing methods that combine AI-assisted extraction with practitioner validation to reduce the gap between what the organization's people know and what its knowledge infrastructure captures. The investment pays returns across every phase of the knowledge lifecycle and across every consumer class, because the quality of downstream knowledge artifacts is bounded by the quality of the acquisition that produced them.

SRE introduced service level objectives and error budgets as quantitative measures of reliability. DevOps introduced deployment frequency, lead time, change failure rate, and mean time to recovery (Forsgren, Humble, and Kim 2018). Knowledge Engineering requires equivalent metrics: coverage (what percentage of organizational knowledge is captured in maintainable form), currency (what percentage of knowledge artifacts accurately reflects the current system state), discoverability (how effectively consumers locate the knowledge they need), transfer effectiveness (whether consumers who access knowledge accomplish their intended task), and maintenance cost (the resources required to keep the knowledge layer trustworthy). These metrics address the invisible infrastructure problem directly: what can be measured can be made visible, and what is visible can be funded. The metrics also provide the feedback loops that a maturing discipline needs to evaluate its own methods and identify where investment produces the highest returns.