Many large organizations now have an AI strategy. It typically sets out the ambition and target position, priority use cases, target architecture, and high-level principles for risk and ethics. Yet execution often remains fragile, and AI initiatives struggle to scale. This pattern can persist even in organizations with strong leadership intent and significant investment.
When AI initiatives lose momentum, the strategy or the algorithm is often not the main cause. The problem usually lies in the organization around them. Who may decide? Who is accountable for the outcome? How is work funded and how lessons from one team reach the rest of the organization?
Many AI strategies focus on what the organization wants to achieve, not on how it must be redesigned to achieve it. They emphasize vision, priority domains, and technology enablers, while treating operating model implications as secondary or downstream concerns. This omission is rarely accidental.
Over the past decade of working on operating models in large industrial organizations, I have seen that operating model redesign is politically sensitive, slow to execute, and difficult to quantify because it touches decision rights, funding authority, performance management, and organizational power. As a result, many AI strategies implicitly assume that execution can happen within existing structures.
This article argues that the operating model is the missing half of AI strategy and sets out what it has to do.
Why AI strategies Lose momentum in traditional structures
The belief that existing structures can carry AI shows up early, as many organizations start by fitting AI into what they already have. Data scientists join functional teams, specialists are assigned, and a center of excellence is added to the existing hierarchy. On paper, this looks pragmatic and barely disruptive. In practice, it leaves the underlying logic of the organization unchanged. 
Traditional organizations are optimized for efficiency, stability, and control. Work is divided into functions, accountability runs vertically, and coordination happens through handovers, approvals, and escalation.
That logic suits predictable work. AI work does not follow it. It cuts across functions, and it creates value through iteration, not delivery. Decisions must be made close to changing data, user behavior and operational context [4].
Put an AI team inside the old structure, and in my experience the tensions surface quickly. Authority is split across functions. Teams answer for outputs but not outcomes. Approval cycles interrupt learning. And when trade-offs arise, nobody has the mandate to resolve them end to end. The structure itself becomes a bottleneck, which is why hiring more specialists is rarely enough. Organizations need forms that let talent work as a system, not as isolated individuals [5].
The central AI team is a phase, not a destination
Some organizations respond by centralizing, often after early experiments have spread across the business. Talent is pooled, standards are set, platforms are shared, and governance feels simpler. For a while, this can work well: early use cases may move faster, technical quality often improves, and leadership gets a single point of contact.
The limits tend to appear as AI moves closer to the core of the business. A central team often sits too far from operations to own outcomes end to end. It builds models and platforms, but not the daily decisions that use them, so adoption depends on persuasion instead of ownership. Demand outgrows capacity, prioritization turns political, and business units start working around the team, recreating the fragmentation it was meant to cure. Above all, accountability blurs, and nobody owns the system from data to decision.
The answer is not to abolish the central team, but to let the model evolve with maturity [5]. Concentrated expertise and consistent standards help when building foundations and learning early. Teams embedded in the business are better placed to scale, because they sit close to context and own the results, but they struggle without shared platforms and standards. For many organizations, the central team is therefore a phase, and its long-term role shifts toward enabling rather than owning.
How the operating model needs to change?
An operating model is how strategy turns into daily work. Who does what, in which flows, with which data and technology, under which rules? In my experience, AI requires several shifts in operating model:
- Funding from projects to capabilities.
- Move decision rights from layered approvals to delegated authority.
- Upgrade governance from upfront gatekeeping to continuous stewardship [1].
- Move from fragmented handovers to persistent ownership and keep one team responsible for the outcome over time.
- Optimize operating model from predictable control to continuous adaptation.
These shifts depend on a shared foundation. Teams need common data, platforms, and standards, so that autonomy does not turn into fragmentation. Data is the clearest case and treats it as a product with an owner, a quality level, and rules for use.
What leaders must do
Leaders have to examine the operating model with the same seriousness as the AI strategy. In practice, three decisions stand out.
Define what the central team owns, what it enables, and what it must let go. Place delivery accountability in the business, supported by the central team. And accept that real control depends more on ownership at the point of impact than on distance.
Three questions can help test whether the operating model is ready. Who decides when the AI is uncertain, and within what limits? Who owns the outcome a year after launch? And where does what is learned in production flow back into the operating model?
These questions matter even more in defence and other high-consequence settings, where the stakes are higher and the constraints tighter. Where an AI system can act, decision limits, accountability, and human oversight must be designed before deployment, not added after the first incident.
Conclusion
An important question for any AI strategy is therefore not only which models to build, but whether the organization is ready to use them. Traditional structures are often well run and serve their original purpose well. But many were designed for predictable work, which may not suit systems that keep learning.
References
- Berente, N., Gu, B., Recker, J., & Santhanam, R. (2021). Managing artificial intelligence. MIS Quarterly, 45(3), 1433–1450. https://doi.org/10.25300/MISQ/2021/16274
- Papagiannidis, E., Mikalef, P., & Conboy, K. (2025). Responsible artificial intelligence governance: A review and research framework. The Journal of Strategic Information Systems, 34(2), 101885. https://doi.org/10.1016/j.jsis.2024.101885
- Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation–augmentation paradox. Academy of Management Review, 46(1), 192–210. https://doi.org/10.5465/amr.2018.0072
- Shollo, A., Hopf, K., Thiess, T., & Müller, O. (2022). Shifting ML value creation mechanisms: A process model of ML value creation. The Journal of Strategic Information Systems, 31(3), 101734. https://doi.org/10.1016/j.jsis.2022.101734
- Fountaine, T., McCarthy, B., & Saleh, T. (2019). Building the AI-powered organization. Harvard Business Review, 97(4), 62–73. https://hbr.org/2019/07/building-the-ai-powered-organization
About the author
Senadin Alisic is Strategic Planning & Development Analyst at Combitech and Saab, working with digital transformation, ecosystems and operating model development. His work focuses on how large-scale organizations, particularly within defence, security, and critical infrastructure, can evolve their operational models and data architectures to bridge the gap between technical AI innovation and real strategic value.
With deep insights into cross-industry ecosystem dynamics, including defence applications at Saab as well as industrial automation cases like smart ports and smart cities, Senadin guides organizations on structuring capabilities for scale, trust, and resilience.
*The views and opinions expressed by the author do not necessarily state or reflect the views or positions of Hyperight.com or any entities they represent.
