A prototype AI assistant helped occasional purchasers follow procurement rules and flagged overlooked requirements in a center-led procurement pilot, reducing instances of maverick buying according to user feedback; the artifact—built on a RAG architecture—offers a practical model for scaling decentralized procurement guidance, but evidence is limited to a four-user pilot.

Main Finding

An AI assistant built on a Retrieval-Augmented Generation (RAG) architecture can effectively support decentralized indirect procurement in a center-led model by acting as (1) a low-threshold first point of contact, (2) a proactive gatekeeper for compliance, and (3) a scalable support mechanism. The assistant helped users surface procurement requirements they might otherwise miss, thereby bridging centralized governance and decentralized execution, improving process compliance, and reducing maverick buying.

Key Points

• Purpose and problem: Support occasional buyers and strengthen compliance in a context of rising procurement volumes with limited centralized resources.
• Methodological approach: Design Science Research (DSR) with an embedded case study; artifact developed through three iterative design-build-evaluate cycles.
• Artifact architecture: Retrieval-Augmented Generation (RAG)
  • Retrieval layer sources organization-specific documents, policies, and historical procurement records.
  • Generative LLM produces user-facing guidance conditioned on retrieved evidence.
  • Mitigation features: provenance citation, escalation/human-in-the-loop, prompts to reduce hallucination.
• Roles the assistant played:
• Low-threshold first contact — reduces friction for occasional buyers to ask procurement questions.
• Proactive gatekeeper — nudges users toward compliant pathways and flags missing requirements.
• Scalable support — lets limited central procurement resources cover more users without linear staff increases.
• Evaluation: Pilot study with four end users using the Technology Acceptance Model (TAM) augmented by trust dimensions; results indicate perceived usefulness, ease of use, and trustworthiness sufficient to support real-world deployment but based on a small sample.
• Design principles produced: provide evidence-backed answers, integrate with existing governance artifacts, enable easy escalation to humans, audit interactions, and iterate on domain-specific retrieval corpora.

Data & Methods

• Research design: Design Science Research (DSR) complemented by a case study in an organizational procurement context.
• Development cycles: Three iterative cycles of design, implementation, and formative evaluation informed artifact refinements (architecture, UI, retrieval corpus).
• Technical implementation: RAG pipeline combining a retrieval index over company procurement policies and documents with a generative model to produce contextualized guidance; provenance linking to sources and fallback to human agents.
• Evaluation:
  • Pilot with four end users (occasional decentralized buyers).
  • Measurement framework: Technology Acceptance Model (perceived usefulness, perceived ease of use, behavioral intention) plus trust-related dimensions (e.g., competence, integrity, benevolence).
  • Outcomes measured qualitatively and via TAM-style responses; pilot showed positive acceptance and examples of prevented non-compliant actions.
• Limitations: small pilot sample, limited duration, and single organizational case — findings are promising but preliminary and context-dependent.

Implications for AI Economics

• Transaction cost reduction: AI assistants lower search and coordination costs for occasional buyers, reducing maverick purchases and the transactions arising from non-compliance.
• Scale economies in procurement services: A RAG-based assistant offers near-linear scaling of governance reach (more users served without proportional staffing increases), improving the productivity of centralized procurement teams.
• Cost-benefit considerations: Up-front investment in building and curating the retrieval corpus, integrating systems, and governance controls is required; expected payback via reduced off-contract spend, fewer compliance failures, and lower support staffing needs.
• Risk management and governance: Improved compliance reduces organizational risk (contractual/regulatory exposure). However, firms must invest in provenance, audit trails, and human-in-the-loop processes to manage LLM errors and trust calibration.
• Market and behavioral effects: Making procurement rules easier to follow can shift buyer behavior and bargaining dynamics (potentially increasing use of preferred suppliers and strengthening center-negotiated leverage).
• Generalizability and adoption: The architectural pattern (RAG + provenance + escalation) is transferable to other organizational guidance tasks (legal, compliance, HR), implying broad potential economic returns where tacit governance knowledge must be distributed cheaply.
• Research & measurement agenda for AI economics:
  • Quantify economic impact (reduction in off-contract spend, time savings, support headcount) with larger-scale randomized or quasi-experimental studies.
  • Model long-run effects on procurement prices and supplier relationships as decentralized buyers shift toward compliant behavior.
  • Study optimal investment levels in retrieval corpus maintenance and human oversight to minimize error costs while maximizing coverage.

Suggested next steps for practitioners: pilot at larger scale with A/B testing against standard support channels, instrument economic KPIs (off-contract spend, time-to-complete procurements, support tickets), and formalize governance (audit logs, update processes for the knowledge base, escalation protocols).