Making the implicit explicit: excavating latent design knowledge with AI

A paper that presents a more accurate neural network for sorting waste, or a better sensor arrangement for predicting when a machine will fail, contains design decisions and reasons for them. The authors rarely say why the design works in a way that could transfer elsewhere, because their field rewards performance, not theorising. The guidance is hard-coded into parameter choices, architecture diagrams, and trade-off discussions.

Design science research has developed careful structures for stating design knowledge, such as the anatomy of a design principle with its aim, mechanism, and rationale. What it lacks is a way to find and assess knowledge that was never stated in those terms. Systematic reviews map topics and summarise findings. Existing tools that index design knowledge assume it is already labelled. Methods for reconstructing design rationale from real-world artifacts exist, but they are manual and slow.

The research question follows: how can latent design knowledge embedded in diverse, non-DSR publications be systematically excavated to formulate reusable design knowledge?

AIDE, short for AI-assisted design knowledge excavation, is framed as an epistemic transformation. The centre column of the framework tracks the state of the team’s knowledge. The left column says what the AI agent does to move it forward. The right column says what the human expert does. At no point does the AI theorise on its own.

The project started in January 2025 from a practical bind. An interdisciplinary team on AI for the circular economy had to analyse nearly a hundred engineering and applied computer science papers for their prescriptive insight, and no method said how. The team bootstrapped heuristics and prompts by experimenting directly in ChatGPT and Gemini, comparing what each extracted. A first attempt to build custom software proved too complex, so the pipeline moved to the workflow tool n8n, with the extraction prompt refined over five rounds on a subset of papers.

Each round was checked from two sides. Circular economy experts judged whether the extracted insight was technically correct and faithful to the source. Design science experts judged completeness, level of abstraction, and fit with established knowledge structures. Six researchers with two to ten years of design science experience and two to seven years in the domain formed the team; one joined late specifically to evaluate the method from outside.

The instantiation is deliberately ordinary: Airtable holds the papers and the human annotations, n8n orchestrates the steps, and Gemini Flash 2.5 does the reading. Ninety-nine papers took an hour and a half. The paper is explicit that these tools are illustrative. The logic of human scoping, AI extraction, and reflexive validation transfers to any stack.

The corpus came from a PRISMA-guided search across Scopus, Web of Science, IEEE Xplore, and the AIS library: 1,116 records, screened to 218 for full-text assessment, and finally 99 papers that reported an implemented artifact, linked AI to a circular economy strategy, and carried enough design signal to reconstruct intent. The papers cover predictive maintenance, automated sorting and disassembly, and optimisation for resource-efficient design.

The worked example is a study on AI in municipal waste management. Its prescriptive intent is implicit: users should be recognised, sensors should measure what they throw away, and incentives should reward good behaviour. AIDE surfaced those fragments and formalised them.

Three implications follow. First, grounding a design project is no longer only about choosing the right theory or collecting your own data. It can mean actively reconstructing prescriptive intent from the long tail of artifact-oriented research in other fields. Second, the paper makes the human–AI division of labour concrete: AI for detection, reconstruction, and schema-conformant drafting at scale; humans for interpretation, abstraction decisions, and validation. Third, literature work shifts from search and synthesis toward excavation and structuring as a core research activity.

The limits are stated plainly. Poorly documented studies offer little to excavate. Language models have context windows, cost, and stability constraints. The demonstration is a proof of concept at the level of single papers; consolidating principles across papers, preserving boundary conditions, and handling competing prescriptions are the next problems, along with comparing the pipeline against manual extraction.