Governments internalize advisory or AI work for three distinct purposes—control, reuse, or brokering expertise—and each purpose demands a different organizational design; mismatched metrics, staffing, or funding predict poor outcomes.

Main Finding

Governments create in‑house consulting units for distinct reasons — control over consequential work, capturing returns from recurring problems, or mobilizing dispersed expertise — and each reason requires a different organizational design. The paper identifies three configurations (Strategic Transformation, Scalable Solutions, Expertise Platform) and shows that viability depends on aligning work, staffing, knowledge, authority, client relationships, financing, and evaluation so the unit can reproduce the conditions needed for future work. Internalization is therefore organization‑building, not merely reversing a procurement decision.

Key Points

• Three motivating logics for a dedicated in‑house unit:
  • Retain control over consequential work (high stakes, strategic projects).
  • Capture returns from problems that recur across government (reuse, efficiency).
  • Mobilize expertise dispersed across public and adjacent institutions (networked know‑how).
• Three corresponding configurations:
  • Strategic Transformation: focus on high‑impact, consequential projects requiring authority and trusted relationships; staffing centers on senior policy/technical leaders; evaluation emphasizes influence and policy outcomes.
  • Scalable Solutions: focus on building reusable products/services for multiple clients; staffing favors engineers and product teams; financing and incentives align to reuse and cost‑recovery; evaluation emphasizes reuse, uptake, and cost savings.
  • Expertise Platform: focus on brokering and aggregating expertise across institutions; staffing emphasizes community managers and boundary spanners; evaluation measures network activation and knowledge flows.
• Design elements must fit together: mismatch (e.g., product metrics for a transformation unit) undermines effectiveness.
• Documentary comparison of eight units in five countries grounds the typology and shows real‑world tradeoffs.
• Evaluation must be configuration‑ and stage‑sensitive: early stages need legitimacy and pipeline development metrics; mature stages emphasize outcomes and reuse.
• Internalization succeeds only if the unit’s work renews the conditions (client demand, authority, funding, knowledge assets) necessary for continued value creation.

Data & Methods

• Comparative qualitative study using documentary analysis of eight in‑house consulting units across five countries.
• Uses case comparison to derive a theoretical typology linking motivations to organizational configurations and design choices.
• Focuses on how work types, staffing models, authority, financing, client relationships, and evaluation practices interact in practice.

Implications for AI Economics

• Choice of in‑house AI strategy must match the government’s objective:
  • If the goal is control over high‑stakes AI (safety, fairness, national security), adopt the Strategic Transformation model: empower the unit with authority, hire senior technical-policy staff, measure policy influence and safety outcomes.
  • If the goal is economy‑wide reuse of AI components (common models, data pipelines), adopt the Scalable Solutions model: product engineering teams, funding for platform development, metrics for reuse/adoption and cost‑savings.
  • If the goal is to pool diffuse AI expertise across agencies and partners, adopt the Expertise Platform model: roles for community facilitation, incentives for knowledge sharing, metrics for network activation and capacity building.
• Misalignment risks: building an AI product shop but charging it with transformational policy work (or vice versa) will likely lead to failure, unsustainable pipelines, or dependence on external vendors.
• Evaluation and funding should be stage‑aware:
  • Early stage: measure pipeline development, trust, and relationship building (inputs/process).
  • Later stage: measure deployment, reuse, impact on public outcomes, and sustained demand (outputs/outcomes).
• Talent and incentives: compensation, career paths, and mixing of secondments vs permanent hires must reflect the configuration; AI talent strategies differ whether you need platform engineers, senior policy technologists, or community builders.
• Governance and IP/data rules: alignment needed between the unit’s mission and data access, IP ownership, procurement rules — especially important for generative models, datasets, and safety tools.
• Research opportunities for AI economics:
  • Empirical evaluation of costs/benefits of different configurations in AI deployment.
  • Measurement frameworks for public value of in‑house AI units across stages.
  • Comparative studies of retention, diffusion, and vendor dependence outcomes under each model.

Short recommendation: Before creating an in‑house AI team, specify which of the three public‑value objectives is primary, design the unit (staffing, authority, financing, evaluation) to fit that objective, and plan stage‑appropriate metrics and governance to ensure the unit can reproduce the conditions for continued work.