Managers and AI: who decides?
Companies are increasingly relying on Artificial Intelligence in decision-making processes.How to exploit the potential of technology without losing the last word?
Michael Schrageis a researcher at the Initiative on the Digital Economy of the MIT Sloan School of ManagementDavid Kironis Editorial Director ofMIT Sloan Management Reviewand Head of the Big Ideas research program
Published in issue 2, March/April 2025 – Mit Sloan Management Review Italy.
The incorporation of Artificial Intelligence (AI) agents at scale transforms how companies define, design and implement decision environments¹. Our research shows that organizations that use AI to generate sets of sophisticated choices – rather than single, 'best' or 'optimal' solutions – achieve superior results across multiple industries. These intelligent systems don't just improve decision making, they push organizations to redesign decision rights, accountability frameworks, and power dynamics among decision makers.
Based on behavioral economics choice architecture principles, our Intelligence Choice Architecture framework captures how these increasingly sophisticated systems reshape corporate decision making² (Box 1). By combining generative and predictive AI capabilities to create, refine, prioritize and present options, ICAs transcend conventional recommendation engines. As AI agents, ICAs can articulate and explain tradeoffs, surface hidden opportunities, and learn from the results to refine future choices. ICAs mark a decisive transition from the use of algorithms, mainly for the automation of tasks, to the use of AI as an architect of superior decision-making environments.
Consider a major retail company whose human resources department employs AI to identify emerging talent in sales and merchandising – a strategic imperative. While the AI system proves adept at identifying high-potential candidates in unexpected corners of the organization, it quickly becomes clear that these exceptional workers will remain undervalued without a new framework of decision rights governing development, transfer and promotion decisions. Success requires implementing a collaborative, data-driven decision architecture that aligns talent development options with management incentives, organizational priorities and concrete outcomes. The lesson is clear: to unleash the value potential of internal talent, you need to reallocate decision-making rights.
As AI capabilities evolve, ICAs will go beyond simple decision support tools and become sophisticated augmentation systems for human decision makers. They will create powerful new frameworks where human judgment and AI work together to increase business value creation (The Research).
Box1 – Intelligent choice architectures (ICA)Intelligent choice architectures are dynamic systems that combine generative and predictive AI capabilities to create, refine, prioritize, and present choices for and with decision makers. They actively learn from the results, seek information, and modify the options available to decision makers.
ResearchThis second article in the Strategic Measurements 2024-2025 series, conducted in collaboration with Tata Consultancy Services, examines how leading organizations integrate predictive and generative AI to develop and present better choices to human decision makers. Based on interviews with executives from six major industry groups, our research reveals the emergence of intelligent choice architectures – a new paradigm in which AI systems proactively participate in structuring and shaping strategic decisions. The implications for organizational performance, decision rights and strategic agility are significant, particularly as companies face increasing complexity and compressed decision cycles.
Decision rights 2.0
The late Harvard Business School professor Michael C. Jensen devoted decades of research to determining how the distribution of decision rights affects corporate performance and what companies can do to allocate them most effectively. Jensen argued that the allocation of decision rights is “an extraordinarily difficult and controversial managerial task,” warning of the potential dangers of hypercentralization and hyperdemocratization of decision making.
As compound AI systems – systems that combine predictive and generative AI – learn to become more sophisticated choice architects, companies' focus shifts from executing decisions to designing them. Managers become responsible for the decision-making environments in which staff members operate, including when AI-generated suggestions should be acted upon and when they can be ignored. How powerful or constraining should the generated intelligent choice sets be for executives' and managers' decisions? Consider, for example, a trading algorithm that discovers a new market pattern. Should it wait for human validation before acting? What about an ICA agent managing supply chain operations and identifying a more efficient logistics strategy: what permissions are needed before implementing it? Under what conditions should the organization encourage human initiative with respect to obedience and compliance? These are questions leaders need to consider.
ICA agents must reflect and respect the values and aspirations of an organisation. In the era of decision rights 2.0, companies must determine who has the authority and responsibility to design, implement and govern the choice environments where human judgment and ICA capabilities intersect. This authority carries explicit responsibility for both immediate outcomes and the long-term effectiveness of decision architectures. This AI-driven redefinition elevates decision rights from a set of business rules and practices regarding who can make specific decisions, what they can decide, and how strategic decisions shape how organizations leverage the combined power of human judgment and AI.
In fact, ICA agents do not simply provide decision support, but create decision-making environments in which the best choices emerge from the interaction between machine intelligence and human judgment. Think of commercial aviation flight management systems that advise pilots: they don't just process navigation data, they store flight path data and adapt to different routes, weather conditions and pilot preferences, all while operating within strict safety parameters. Likewise, corporate ICA agents continually learn, operating within clear operational, legal and regulatory boundaries.
This directly addresses the all-too-common fear that increasingly intelligent and capable AI systems will render human judgment marginal and/or irrelevant. In reality, the opposite is true. When ICA agents take on the heavy lifting of data analysis, pattern recognition and optimization, they free up their human counterparts and collaborators to focus on higher-order challenges.
Liberty Mutual has actually created an ICA agent to help train new claims adjusters, offering them more personalized training based on 20 thousand articles of company knowledge. This Ica agent helps adjusters more efficiently handle incoming customer calls to resolve claims quickly. The Ica agent is one of the implementations of AI Gen across the company. Additionally, a year after implementing LibertyGpt, OpenAI's internal instance of ChatGpt, company-wide, Liberty Mutual has seen internal employee productivity improve and sustain. Says Monica Caldas, the company's Global Chief Information Officer, that it has saved more than 200,000 hours per person compared to previous settlement workloads.
With ICAs, important business decisions depend on the nature and purpose of intelligent decision environments as much as markets, products, culture or strategy. A new focus is emerging on metadecision rights: the design and governance of the systems that generate choices. A new metadecision-making imperative requires human leadership teams and intelligent algorithms to come together to determine how decision rights around decision rights should be effectively allocated by human leaders and the cutting-edge algorithms responsible for those rights. Ironically, leaders seeking to maximize the value of AI have little choice but to satisfy these metadecision obligations.
Seismic shifts in the agency Artificial Intelligence enterprise
We see three structural shifts underway that will require leaders to proactively address the allocation of decision rights, power, and decision-making practices in the agency AI enterprise.
Energy flows for architects of intelligent choices.As ICAs spread throughout the enterprise, the focus of decision rights will shift from those who make decisions to those who design better decision environments. This is the case of the French multinational pharmaceutical company Sanofi, where ICAs systematically contribute to improving the decision-making process. The company's research and innovation teams have adopted data-driven strategies to optimize project investments and overcome challenges such assunk-costbias, a common trap in the industry that can make it difficult to transition from underperforming businesses. Sanofi's AI systems now enable leaders to confidently redirect resources based on data-driven insights. This transformation of decision making demonstrates the power of well-designed ICAs to amplify human judgment and ensure decisions are aligned with business objectives. As Emmanuel Frenehard, Chief Digital Officer at Sanofi, observes: “We use ICA first and foremost to systematically reduce human cognitive biases.” Ensuring that portfolio decisions are driven by data rather than emotional attachments or past spending is a key function of Sanofi's ICAs.
Network effects amplify and enhance decision-making intelligence.ICAs create network effects whereby each decision improves the predictive accuracy of the system and the quality of future decisions. Payment technology multinational Mastercard's ICAs, such as those that proactively identify transaction patterns to prevent fraud and address unnecessary card declines, exemplify this virtuous cycle. Every transaction that passes through the system enriches the effectiveness of ICA, leading to better anomaly detection and a better customer experience. Leaders design decision-making environments to continuously refine themselves as they work, fueling network effects. This ensures that ICAs' learning capabilities benefit the entire organization (not to mention merchants, card issuers and cardholders). Greg Ulrich, Chief Data and Artificial Intelligence Officer, explains: "With more data, we add more value to our services. With more services, we add value to our payments ecosystem."
Real-time optimizations redefine authority and oversight.Algorithmic trading and programmatic advertising illustrate how ICA agents can autonomously and dynamically refine tactics to maximize results at a faster rate than human managers can achieve. In these contexts, handing over decision-making authority to these agents increases the likelihood of achieving positive outcomes. This phenomenon is also observed in revenue forecasting. In a pharmaceutical company, an ICA agent was used to generate revenue forecasting alternatives. As the model evolved, more accurate forecasts were generated by collaboration between financial managers and the ICA agent. In other words, the ICA agent has transformed from a decision support tool to a collaborator. Ultimately, the Ica agent became even more precise without human intervention. In this context, the quasi-autonomous agent has gained the authority to define revenue projections and dynamically modify them based on the emergence of new information.
These three changes require a fundamental rethink of corporate power structures. Success in the age of AI belongs to human executives who are willing to collaboratively build and orchestrate intelligent choice environments that ensure human judgment remains adequately engaged in creating value. The question for leaders charged with organizational design will shift from “Who gets to decide?” to “How can we design better ways to decide?”.
Anticipate power conflicts
Pragmatically, ICAs are not limited to being tools that inform human decisions, but are entities that can learn to develop decision-making capabilities on their own. These capabilities represent a fundamental governance challenge: as ICAs learn and improve, they transform from simple decision support tools into increasingly capable decision makers, who can match or even surpass expert human performance. When an ICA's ability to make effective decisions exceeds its formally granted decision rights, an inherent conflict is created: the learning-authority dilemma. To solve this problem and harness the full potential of these intelligent systems, organizations need dynamic governance frameworks that systematically evaluate the capabilities of ICAs and intentionally expand their authority when warranted, while maintaining adequate oversight and controls. Balancing the benefits of enhanced capabilities with maintaining accountable governance as ICAs evolve beyond their initial decision constraints will become a leadership imperative for deploying AI agents at scale (Box 2).
Box 2 – Decision-making rights in the AI eraOrganisations must rebuild decision rights frameworks around three fundamental principles:1.Architectural authority: The advantage will be for those who know how to best design the environments of choice.2.Network intelligence: Decision rights will extend to all human-AI networks.3.Dynamic accountability: Measurement systems will evaluate both choices and outcomes.
The emergence of meta-responsibility
As AI becomes more sophisticated at designing choices for and with humans, the most critical decision rights will focus on metadecisions, or choosing how to design the systems that make the choices. A new hierarchy of decision rights will thus be created, in which the power to shape decision environments will replace the power to make individual decisions. Leaders are becoming accountable not only for the decisions they make, but also for the quality of the ICAs they create.
Among the actions leaders can take to move in this new direction are:
Anticipate governance for AI-driven choice systems: companies must establish ethical, strategic and operational boundaries for their ICAs.Elevating cognitive contributions: This requires a shift from tactically enabling decision making to overseeing the design of intelligent decision environments.Ensure alignment: Leaders must create systems and processes to ensure that AI-generated decisions advance the organization's values and goals.Establish new metrics: these should measure the quality and diversity of choices, the effectiveness of the decision-making environment, learning and adaptation rates and the optimization of network intelligence.
The future does not belong to those who make the best decisions, but to those who create the best decision-making environments.
We would like to thank the following people, who were interviewed for this article:Monica Caldas, Global CIO, Liberty Mutual InsuranceEmmanuel Frenehard, head of digital at SanofiGreg Ulrich, head of data and AI at Mastercard
NOTE:
1. In the context of this research, decision environments are dynamic environments in which people and algorithms consider complex information, competing priorities, and constraints to make consequential decisions.
2. M. Schrage and D. Kiron, “Intelligent Choices Reshape Decision-Making and Productivity,” MIT Sloan Management Review, October 29, 2024,https://sloanreview.mit.edu.
Show all references.
3. Jensen, M. “Decision Rights: Who Gives the Green Light?”. Harvard Business School Working Knowledge, August 8, 2005,www.library.hbs.edu.