Rather than chasing shiny AI tools, small and mid-sized firms should rebuild data foundations: lightweight, modular process ontologies and cross-functional mapping via knowledge graphs deliver faster, lower-risk returns on AI investments. This incremental approach promises better interoperability and more sustainable productivity gains than large, disruptive data overhauls.

Main Finding

The most disruptive and cost‑effective step for small-to-medium enterprises to become AI‑driven is to (re)build lightweight, modular data foundations focused on processes and workflows rather than chasing pointy, high‑profile AI tools. Knowledge graph techniques—designed as minimal, modular, process‑aware, and cross‑functionally aligned artifacts—are central enablers of this low‑impact, high‑return data strategy.

Key Points

• Counter‑intuitive recommendation: "step back" from hype and invest in foundational data work (metadata, lineage, ontologies) to make downstream AI robust and sustainable.
• Executive buy‑in requires linking foundational work to business KPIs (ROI, Time‑to‑Market, reduced failure rates), not just generic claims about "better data = better AI."
• Data Mesh principles (domain ownership, distributed governance) provide an organisational framework that aligns with knowledge graph adoption.
• Three practical ontology priorities:
  • Model processes and workflows (dynamic sources of change), not just static data relationships.
  • Use modular, minimal ontologies aligned to domains to minimise overhead and support rapid prototyping.
  • Advance ontology alignment to map cross‑department glossaries and workflows rather than imposing a single vocabulary.
• Technical enablers discussed: semantic lifting of workflows (e.g., BPMN→OWL), Large Language Models for ontology learning and semantic extraction, modular networks of knowledge graphs, and alignment/negotiation techniques for resolving cross‑functional disagreements.
• Recommended approach: incremental, minimally invasive interventions that automate capture of knowledge, enable routine updates, and prioritise high‑value processes.

Data & Methods

• Type of paper: vision/position statement (conceptual synthesis), not an empirical study.
• Methods used:
  • Literature synthesis and discussion of recent industry patterns (cites Data Mesh literature, ontology matching/learning, semantic lifting, LLM tools).
  • Practitioner/author experience and interpretation of industry pain points (start‑up → scale‑up transition problems).
  • Technical pointers to existing research and prototypes (e.g., BPMN formalisation, LLM‑based ontology learning, ontology negotiation literature).
• No original datasets, experiments, or quantitative evaluation presented; recommendations are normative and grounded in prior work and observed industry practice.

Implications for AI Economics

• Firm-level productivity and returns:
  • Better data foundations reduce the probability of AI project failure and can materially improve ROI on AI investments by improving model reliability and reducing rework.
  • Faster, reliable time‑to‑market through interoperable, process‑aware metadata that speeds integration across product/engineering/business units.
• Cost structure and investment timing:
  • Advocates for smaller, staged investments (lower upfront fixed cost) that can be amortised and scaled—reduces financial risk for SMEs.
  • Emergence of a market for lightweight ontology/knowledge‑graph tooling and services (semantic lifting, modular KG platforms, alignment tools).
• Organizational and transaction costs:
  • Domain‑owned, modular ontologies reduce coordination costs within firms (aligns with Data Mesh), but require governance investments and possibly new roles (semantic engineers, ontology stewards).
  • Cross‑functional mapping (rather than enforced standardisation) can lower resistance to change and reduce bargaining/coordination frictions.
• Labor and skills:
  • Demand shift toward roles that combine domain knowledge and semantic modelling / data governance skills; potential for reskilling rather than wholesale automation of knowledge work.
• Competition and heterogeneity:
  • Firms that adopt these foundations gain sustainable advantage in deploying AI reliably; the approach may increase heterogeneity in AI outcomes across firms (those who invest foundationally vs. those who chase tools).
• Policy and measurement:
  • Necessitates new KPI linkage and empirical measurement of how foundational data work affects economic outcomes (ROI, TTM, error rates)—opening avenues for empirical AI‑economics research.
  • Regulators and auditors may value process‑aware metadata and lineage for compliance and accountability, affecting compliance costs and market trust.
• Technology diffusion:
  • Advances in LLMs and tooling that lower the cost of semantic lifting can accelerate adoption, but economic gains depend on organisational uptake and governance practices.