Michael Majster, Partner at Arthur D. Little – Interview Series

Michael Majster, Partner at Arthur D. Little, is a seasoned strategy and technology consultant with deep expertise in digital transformation, IT strategy, and innovation-led growth. Based in Brussels and working across the Benelux region, he brings more than two decades of experience advising CIOs and senior executives on driving large-scale change and delivering measurable business outcomes. Before joining Arthur D. Little, he spent 16 years at Accenture, where he held senior leadership roles, including Managing Director for Technology Resources in France and the Benelux, and developed strong sector expertise across energy, utilities, chemicals, and natural resources.

Arthur D. Little is one of the world’s oldest management consulting firms, founded over 135 years ago and widely recognized for its focus on innovation, strategy, and technology-driven transformation. The firm works with global organizations to solve complex business challenges, combining strategic insight with deep industry expertise to support areas such as digital transformation, innovation management, and operational improvement. With a strong emphasis on bridging strategy, technology, and execution, Arthur D. Little has built a reputation for helping clients develop breakthrough innovations and adapt to rapidly evolving market dynamics.

You spent over 16 years at Accenture, ultimately leading technology and resources across BeNeLux and France, before becoming a Partner at Arthur D. Little. How did that journey shape your perspective on what it really takes for organizations to move from AI experimentation to large-scale operational transformation?

During my time at Accenture, I saw the company expand from 30,000 to over 500,000 employees. I experienced firsthand how IT delivery can be industrialized to an extreme degree to support large-scale transformations for CIOs.

Over the years, however, IT has increasingly become a core part of the business, driving a need for organizations to move away from treating it as a commodity.

While moving to Arthur D. Little, I wanted to explore how these strategic transformations of core business could be dealt with from within, how bespoke solutions instead of standardized platforms would better enable differentiation, and joint accountability of the whole ExCo could help organization make a step change in value creation.

In your recent viewpoint, It’s not the agents, it’s the system, you argue that multi-agent AI is less about technology and more about operating model change. Why do you think so many organizations still approach AI primarily as a tooling problem?

Organizations tend to approach AI as a tooling problem because it fits neatly into existing structures, budgets, and ownership, avoiding the complexity of cross-functional change. This enables quick pilots and visible progress, but often results in initiatives that fail to scale or deliver meaningful ROI. Many leaders also rely on legacy software mental models, assuming AI can be deployed like traditional IT systems rather than requiring workflow and decision redesign. At the same time, accountability for operating model change is diffuse and harder to manage, making technology-led initiatives more attractive. As a result, companies default to tools because it is easier and safer than confronting the deeper transformation needed to unlock AI’s full value.

Many enterprises are currently experimenting with agentic AI but struggling to scale. What are the most common structural barriers preventing organizations from moving beyond pilots into production?

Organizations struggle to scale agentic AI because they focus on isolated use cases rather than redesigning end-to-end processes with clear ownership. Operating models are often not adapted, with unclear roles, decision rights, and insufficient integration of human–agent collaboration into daily workflows. Many initiatives also lack strong links to measurable business outcomes, causing them to stall in pilot phases without demonstrating ROI. In addition, fragmented data and legacy systems limit the ability of agents to operate seamlessly across functions. Finally, weak change management and unresolved governance concerns around risk and reliability further slow adoption and prevent scaling.

You emphasize that orchestration, not individual agents, is where the real value lies. What does effective orchestration actually look like in a real enterprise environment?

Effective orchestration means designing AI around end-to-end business processes, not individual tools, with agents assigned clear roles across the full workflow. Multiple agents collaborate like a team-handling intake, decision-making, execution, and supervision – coordinated through a central orchestration layer. These agents are tightly integrated with enterprise systems via APIs and data sources, enabling real actions rather than just insights. The orchestration layer also manages task sequencing, exceptions, and escalation to humans, ensuring smooth and controlled execution. Finally, strong governance, validation mechanisms, and feedback loops are embedded to ensure reliability and continuous improvement.

As organizations redesign workflows around AI, how should decision rights evolve between humans and autonomous systems to avoid both over-automation and bottlenecks?

Decision rights should evolve into a risk-based model where low-risk, high-volume decisions are automated, while high-impact or uncertain cases remain under human control. Agents should handle perception, analysis, and execution, while humans focus on judgment, exceptions, and defining correct outcomes. Clear escalation pathways are essential so agents know when to defer decisions rather than overstepping their authority. To avoid bottlenecks, human involvement should be selective and focused on exceptions instead of full oversight of every step. At the same time, strong guardrails, validation mechanisms, and monitoring ensure control and prevent over-automation.

One of your key points is that reliability must be engineered end-to-end. What are the biggest failure modes you are seeing today when companies deploy multi-agent systems without sufficient safeguards?

A key failure mode is error propagation, where small mistakes by one agent cascade through the system and become amplified. Another issue is hallucination and false confidence, as AI can generate plausible but incorrect outputs without signaling uncertainty. Organizations also struggle when too much autonomy is given to agents in high-impact decisions without proper validation or approval gates. Weak exception handling further compounds the problem, with agents failing to escalate ambiguous or edge cases to humans. Overall, failures occur because reliability is not engineered across the full system of coordination, verification, and oversight .

Governance is often treated as a compliance layer rather than a design principle. How should companies rethink governance when building AI-native operating models?

Governance should shift from a post-hoc compliance layer to a core design principle embedded directly into AI workflows. This means integrating guardrails, validation steps, and approval gates at key decision points rather than adding controls after deployment . Companies should adopt a risk-based approach, adjusting levels of autonomy and human oversight depending on decision impact and uncertainty. Clear accountability and decision rights must be defined, ensuring agents know when to act and when to escalate. Finally, continuous monitoring, feedback loops, and engineered reliability mechanisms are essential to maintain trust, performance, and control at scale.

From your work advising CIOs and CDOs, what separates organizations that achieve measurable ROI from AI versus those that remain stuck in perpetual experimentation?

Organizations that achieve ROI focus on end-to-end process transformation rather than isolated AI use cases. They link initiatives to clear financial KPIs and embed AI into workflows, roles, and decision-making structures. They also invest in change management, adoption, and governance to ensure solutions scale effectively. In contrast, organizations stuck in experimentation run disconnected pilots, focus on tools instead of processes, and lack clear ownership and accountability. As a result, they fail to translate AI into measurable business value and remain trapped in the pilot phase.

As multi-agent systems become more complex, how should organizations think about observability, monitoring, and escalation to maintain trust in AI-driven decisions?

Organizations should treat observability, monitoring, and escalation as core system design elements, not afterthoughts. This includes ensuring full traceability through logging, audit trails, and validation checkpoints so decisions can be understood and verified . Monitoring should focus not only on performance but also on outcome quality, errors, and how issues propagate across agents. Clear escalation rules are needed so agents defer to humans in high-risk or ambiguous situations without creating bottlenecks. Ultimately, trust comes from systems that are transparent, measurable, and designed for human oversight by default.

Looking ahead, how do you see enterprise operating models evolving over the next 3 to 5 years as multi-agent AI systems become more deeply embedded into core business processes?

Enterprise operating models will evolve toward hybrid human–AI workforces, where agents act as digital employees embedded in core processes. Organizations will shift from functional silos to process-centric structures with end-to-end ownership and orchestration across workflows. Decision-making will become more dynamic, with routine tasks automated and humans focusing on exceptions and oversight. Governance and reliability mechanisms will be built into daily operations rather than treated as separate layers. Ultimately, competitive advantage will depend on how well companies redesign their operating models around AI, not just on the technology itself.

Thank you for the great interview, readers who wish to learn more should visit Arthur D. Little.