From Automation to Agency: Leading in the Era of Agentic AI

AI agents are as transformative as the advent of the internet. They will change how we organize work, manage operations, and drive value

A question I often hear from AWS customer executives is how they should think about leading in this new era. I use the same mental models I use to lead my most independent, high-agency employees. These employees assess situations, make judgment calls, and deliver results based on their understanding of strategic intent. AI agents operate similarly but at machine scale.

As executive leaders, we already know how to run complex operations: set clear objectives, establish boundaries, and measure outcomes. But with AI agents, we’re entering new territory. Unlike traditional systems that follow precise, predefined instructions and are expected to behave consistently, AI agents are nondeterministic, adapting their approach based on context and learning from each interaction.

Think of it as applying time-tested leadership principles to a new kind of team member, one that combines humanlike decision-making with machine-scale actions. Our challenge is to develop governance, risk management, and operational models that embrace this fundamental difference.

After working with hundreds of AWS customer executives, here are some mental models that I find helpful.

Governance: From Direct Management to The Board of Directors

Think about how your board of directors interacts with you. They don’t manage your daily decisions. They align on strategy, define success metrics, and maintain oversight. Between board meetings you operate independently, making decisions based on your understanding of the company’s direction and risk boundaries.

AI agents function similarly. They make autonomous decisions based on the strategic context we provide, but don’t check with us on every choice. A successful board doesn’t micromanage its CEO; it sets a clear direction and offers intelligent oversight. We don’t need to micromanage our AI agents either.

First consider strategic direction. Your board tells you where you’re heading, not exactly how you get there. When working with AI agents, provide strategic intent and expected outcomes, not procedural steps.

Establish clear decision-making boundaries. Your board establishes what decisions require their approval and what falls within executive authority. Do the same and define the scope of your AI agents’ autonomy while establishing boundaries and protocols for when they must escalate decisions.

Don’t forget periodic recalibration. Board meetings assess overall performance, ensure strategic alignment, and adjust direction as needed. Our approach to AI agents should do the same. Establish regular reviews to evaluate their effectiveness, refine their decision-making frameworks, and ensure their alignment with strategic goals.

Risk Management: From Factory Floor to Trading Floor

Traditional risk management operates like a factory floor. It’s predictable and controlled, with clear rules and procedures. Agentic AI risk management operates like a trading floor. Traders have real-time authority to make decisions within defined parameters, while the firm maintains oversight.

Modern trading floors use sophisticated real-time risk monitoring. These systems track exposure and flag unusual patterns instantly. AI agents require similar vigilance, but with added complexity. We need to detect when their behavior drifts from expected parameters and when their cumulative actions create emergent risks that weren’t apparent in isolated decisions.

Market circuit breakers—automatic safeguards that halt trading during extreme volatility—are another useful parallel. AI systems need a similar mechanism but with more nuanced triggers. We should be able to pause operations for clear-cut threshold breaches as well as subtle pattern deviations that could signal unforeseen risk scenarios.

Position limits also translate well to AI risk management. Traders can’t exceed certain risk thresholds without approval; AI agents need similar constraints but with adaptive boundaries. We don’t need to prescribe every action, but we should establish clear risk boundaries that agents cannot cross without human intervention.

Organizational Impact: From Functional Silos to Immune Systems

Most modern organizations operate as interconnected but separate departments, each with its own specialized function and boundaries. AI-enabled enterprises will function more like an immune system, with distributed intelligence operating across the organization, responding to challenges anywhere in the system, and continuously adapting based on what it learns.

Consider how this changes our approach to cross-functional work. AI agents transcend traditional organizational boundaries, creating value by connecting work across silos. We’ve seen this before with other major technological evolutions. Cloud computing wasn’t just an infrastructure change; it collapsed the boundaries between our development and infrastructure teams. But AI agents will drive an even more fundamental shift. Their impact will extend far beyond IT, breaking down barriers across all business functions, from finance to marketing to operations to customer service.

This change also impacts how we think about business processes. Today we think in terms of step-by-step workflows, like a predetermined sequence of handoffs. It’s like a relay race where each runner knows where to pass the baton. AI agents work differently. They understand objectives and context, then dynamically orchestrate responses based on how situations unfold. We’ve seen this pattern before. ERPs didn’t simply digitize our existing processes. They forced us to fundamentally reengineer workflows across finance, marketing, HR, and operations. AI agents represent the next evolution; we need to reimagine linear workflows as dynamic, context-aware processes that adapt in real time.

Perhaps most significantly, AI agents transform how organizations learn and retain knowledge. Organizations today struggle with institutional memory. Knowledge stays trapped in departmental silos; lessons fade as people move on; and context gets lost between handoffs. AI agents retain and build upon every interaction, continuously synthesizing insights across domains and improving systems. When one agent discovers a better approach, that learning becomes immediately available across the network.

Culture: From Operational Execution to Continuous Learning

The most profound change AI agents demand isn’t technological—it’s cultural. Most organizations optimize for consistency and predictability. We value standardized processes and repeatable outcomes. Leaders are typically rewarded for flawless execution of predetermined plans. But AI agents require us to adopt a different cultural mindset, one that embraces adaptation and evolution based on continuous learning.

Research laboratories are instructive models. They combine systematic methods with an openness to unexpected discoveries. Researchers succeed when they rapidly learn and adapt based on evidence. We need this balance of structure and flexibility to deploy AI agents effectively.

In this new cultural model, our approach to operations becomes more dynamic. Instead of relying on fixed processes, cultivate an environment that encourages exploration and iteration. Encourage your teams (both human and AI) to discover novel approaches, take different paths to solve the same problem, and be curious about capturing the lessons. Your goal isn’t the perfect execution of predefined steps. It’s the continuous discovery of better ways to achieve desired outcomes.

This shift to a learning culture changes how humans and AI agents interact. We move from being process operators to learning partners. We can comfortably question an agent’s recommendations, understand its reasoning, and work together like good colleagues to refine our approaches.

We also build a culture of rapid feedback loops to capture and analyze the final outcomes and learn from near misses. As leaders, we play a vital role in this change, modeling curiosity, encouraging calculated risk-taking, and valuing learning as much as immediate results.

Putting Models into Practice

The principles we’ve explored come together when applied to manage outcomes rather than fragments of a workflow.

Consider customer support. Instead of having AI agents handle some parts of every support request, a more effective approach is to let them manage specific issues end-to-end. For example, an agent might fully handle all billing inquiries, from initial contact through resolution. This creates clear boundaries while allowing for autonomy. The agent can access customer history, billing systems, and authentication tools to orchestrate the entire resolution journey.

The governance model guides this journey through clear objectives, like customer satisfaction targets and resolution times. Risk controls mirror trading desk principles: “Escalate to humans if customer sentiment drops below 70%,” or “Pause operations if resolution time exceeds 10 minutes,” or “Flag for review if the solution drifts more than X% from past patterns.” Like an immune system, the agent can access data and capabilities across organizational boundaries to craft comprehensive solutions. And each interaction improves future responses through continuous feedback and adaptation to reflect our learning culture.

As leaders, our opportunity lies in reimagining end-to-end processes with AI agents as integral team members. We can apply the leadership principles we know to this new era of autonomous intelligence.