Giving employees AI access rarely produces true transformation; a five-part socio-technical architecture (AX-5R) that aligns readiness, workflow redesign, decision rights, risk controls and measurable returns significantly improves artifact quality in a controlled ablation study, suggesting firms must redesign work systems—not just distribute tools—to capture AI value.

Main Finding

Adoption of generative AI rarely becomes AI transformation (AX) because firms insert tools into existing routines without redesigning the socio-technical system. The paper proposes AX-5R, a minimum socio-technical architecture—readiness, redesign, role, risk, return—that converts fragmented AI use into accountable, governable, and measurable work systems. In a controlled ablation probe, artifacts produced under the full AX-5R frame significantly outperformed a sham five-part control and four ablations, supporting the claim that configurational alignment of social and technical elements improves transformation outcomes.

Key Points

• Adoption vs. transformation: granting tool access is necessary but insufficient; transformation requires reconfiguring workflows, decision rights, oversight, and measurement.
• Socio-technical framing: failures are not primarily productivity/technology defects but mismatches between AI tools and organizational task interdependencies, accountability, and feedback mechanisms.
• AX-5R architecture:
  • Readiness — set boundaries and prerequisites for AI work.
  • Redesign — restructure tasks and interdependencies.
  • Role — assign accountability and decision rights.
  • Risk — establish controls, governance, and oversight loops.
  • Return — define metrics and feedback to measure value and enable learning.
• Configurational alignment: AX is a joint optimization problem of technical and social subsystems; all five functions are interdependent and must be aligned.
• Practical artifacts: the paper supplies a failure-mode derivation matrix, implementation artifacts, and six testable propositions to guide deployment and evaluation.

Data & Methods

• Empirical probe:
  • Generated 252 artifacts across 3 distinct workflows.
  • Seven prompt arms (including a sham five-part control) implemented across two language models.
  • Blind cross-provider judges scored artifacts; an independent human-expert-rating check validated results.
  • Statistical testing: full AX-5R frame outperformed the control and four artifact-relevant ablations; results significant under two-sided Holm-corrected testing.
• Approach: synthesis of seven literature streams to map common failure modes to the five design functions and derive a minimum socio-technical architecture.
• Deliverables: failure-mode matrix linking problems to design responses, concrete implementation artifacts (templates, roles, control constructs), and six propositions for future testing.

Implications for AI Economics

• Measurement and returns: AX-5R highlights that realized productivity gains depend on organizational redesign and measurement systems (the "return" component). Economists should treat raw tool adoption as an input, not a proxy for transformation or realized output gains.
• Diffusion and adoption puzzles: the framework explains why diffusion of AI tools does not automatically yield macro productivity effects—coordination frictions, missing accountability, and absent feedback loops block value capture.
• Complementarities and misalignment: AX is characterized by complementarities between technology and organizational design. Ignoring these leads to underestimation of required non-AI investments (training, governance, process redesign), changing effective costs and adoption thresholds.
• Labor and firm-level effects: redesign and role assignment reshape task boundaries and decision rights, with implications for skill demand, job content, and distribution of rents within firms. Governance and risk controls affect how work is delegated to AI versus humans.
• Policy and regulation: regulators and policymakers should look beyond tool access and consider incentives, disclosure, and standards that encourage firms to implement full socio-technical transformations (e.g., requiring performance measurement, accountability regimes).
• Research directions: need for longitudinal and field experiments to measure persistence of AX gains, heterogeneity across industries/workflows, cost of reconfiguration, and equilibrium effects on wages, entry, and productivity aggregation.