Organisational practices, not algorithms, decide whether in-house AI pays off: a Finnish telecom’s experience shows alignment, last‑mile integration, governance and a product mindset explain who captures financial gains, while weak change management and misaligned metrics block value.

Main Finding

The gap between AI/ML capability and realized monetary benefit is primarily organizational rather than technical. In a Finnish telecom case, five organizational drivers explain variance in value realization: AI maturity over time, strategic alignment, triple‑level last‑mile integration (process, people, systems), governance, and financial metrics validated by controllers. Failures stem from misaligned process reengineering and weak change management; success is greatest where AI is treated as an ongoing product supported by historical data leadership. A governance tension emerges: the data function prefers decommissioning models while business units prefer retaining them.

Key Points

• Primary insight: organizational structures, incentives and practices determine whether in-house AI converts into financial value.
• Five drivers of realized value:
• AI maturity over time — sustained development and learning leading to predictable outcomes.
• Alignment — strategic and operational fit between AI initiatives and business goals.
• Triple‑level last‑mile integration — coordinated integration across processes, people (skills & roles), and systems at the point where model outputs are used.
• Governance — decision rights, lifecycle management and cross‑unit coordination.
• Financial metrics validated by controllers — credible, controller‑approved measures of benefit that enable funding and accountability.
• Key obstacles: an ill-aligned process reengineering program and inadequate change management for people and roles.
• Successful cases treat AI as a product (ongoing ownership, maintenance, roadmap) and emphasize historical-data leadership (data stewardship and continuity).
• Governance conundrum: central data teams may favor decommissioning for hygiene/cost reasons, while business units prefer retention for practical benefits — this misalignment creates value loss.

Data & Methods

• Case: single organizational case study at a Finnish telecommunications firm.
• Data: ten semi‑structured interviews with stakeholders and internal documentary evidence.
• Analysis: applied thematic analysis and the Gioia method for data structuring and concept generation.
• Theoretical lenses: Resource‑Based View, Dynamic Capabilities, and AI Value Realisation frameworks to interpret organizational drivers and mechanisms.
• Limitations: single‑case, qualitative design limits generalizability; nevertheless provides rich, actionable insights into organizational mechanisms.

Implications for AI Economics

• Valuation & ROI: Financial benefit of AI cannot be inferred from technical performance alone; economic assessments must incorporate organizational readiness, governance costs, and change management investments.
• Investment decision-making: Controllers and finance functions need validated, standardized metrics to credibly evaluate AI projects and enable capital allocation; absence of such metrics leads to under‑ or mis‑investment.
• Incentive design: Align incentives across data teams and business units to avoid premature decommissioning or orphaned models; consider contracting/chargeback, SLAs, or product‑owner roles to internalize benefits and costs.
• Lifecycle accounting: Treat AI models as managed products with explicit lifecycle costs (development, maintenance, monitoring, decommissioning) to improve accounting, amortization and cost‑benefit calculations.
• Organizational capability building: Economic returns depend on dynamic capabilities (continuous learning, cross‑functional integration). Policy or firm‑level interventions that build data leadership and product mindsets increase expected returns.
• Governance tradeoffs: Centralization (for hygiene/cost control) vs decentralization (for business retention and agility) have distinct economic implications; governance designs should balance model risk, maintenance costs, and captured value.
• Evaluation frameworks: Cost‑benefit analyses should include nontechnical frictions (change management, process redesign misalignment) and measure realized rather than projected benefits.
• Generalizability caution: Empirical conclusions are context‑dependent; further quantitative or multi‑case work is needed to estimate effect sizes for use in macro or firm‑level AI economic models.