What racing can teach us about agentic AI and governance

An AI agent is a lot like a race car driver. It has autonomy and is empowered to decide based on its objectives, environment and obstacles. The driver’s success depends on meticulous planning, real-time decision-making and continuous improvement, from the car's aerodynamic design to the strategy behind pit stops. Similarly, evaluating, monitoring, as well as protecting data and AI are paramount for enterprises looking to scale and grow effectively.

At IBM®, AI has helped the company save over USD 3.5 billion. When employees adopt self-service options for routine tasks and manual HR tasks are reduced, it leads to increased capacity across an enterprise. The increased capacity can result in 50% to 60% savings in HR service delivery costs. 

There are now AI agents that can autonomously adapt to new data, learn from (or reinforce) their mistakes and correct decisions to align with their intended purpose. 
Race cars have technology and features built in to protect the drivers, the fans and the infrastructure. But what protects AI agents, the people they interact with, and the data and organizations within which they operate?  

Almost every organization is thinking about how to do more and be more efficient. According to Nielsen Norman Group, agentic AI can be used to augment and increase productivity by 66%, allowing people to focus on what makes the difference in their daily work. However, developing an agent that is proficient, efficient and responsible takes significant effort and planning.

Evaluation takes a cross-functional team effort, from individuals up to the C-suite, across AI, data, compliance, security, risk and privacy . A team effort is necessary to understand how the agent performs in various circumstances, which models work best and what resources and tools the agent should use. Furthermore, the team must evaluate performance during the build and experimentation stage and monitor once deployed.

Agents must navigate a complex data, tool and model landscape, much like a professional driver navigates a complex racetrack. Drivers aren’t alone when racing, and agents of the future are going to interact with many other agents to complete their tasks. Racing requires considerable support: the pit crew and manager; established rules and restrictions of the race; and the technology including the car safety features, components and physical guardrail barriers.

The pit crew ensures that the car is in optimal condition, makes strategic decisions during the race and adapts to changing conditions. At the same time, the other team members check their performance and make decisions based on observed data.

How does this play out in AI? Governance involves managing the entire lifecycle of an AI agent, from development to deployment to retirement. It assesses and manages various relevant risks, ensuring the agent and underlying AI remain accurate, fair and compliant with regulations. This process is similar to a pit crew that ensures the car remains in top condition.

Real-time data and analytics are also crucial for strategic decision-making in racing. There’s no time to go back to the lab to wait for analysis. Continuous monitoring and evaluation are essential for effective lifecycle governance in AI. This process involves tracking model performance, identifying drift and making necessary adjustments. These capabilities ensure that the AI model remains effective and reliable, like a race car staying competitive throughout the race.

A race car must adhere to stringent rules and safety standards; AI has to operate within clearly defined ethical boundaries. Here AI governance serves as the chassis, providing a set of robust principles and regulations to build upon that guide AI's behavior and ensure fairness, transparency and accountability.

Its mechanisms for monitoring and auditing AI systems are like tires, ensuring alignment with governance principles and legal requirements. These safeguards prevent unwanted drifts into biased decision-making or privacy violations, much like pit stops ensure a race car stays on course and performs at peak capacity.

In this high-stakes race, stakeholders from all angles—developers, assurance and security professionals, business leaders and regulators—must work in tight formation, constantly communicating and adjusting to maintain control over the AI's trajectory. Some of the changes we considered unpredictable, can now be predicted. This role belongs to the director and engineers, which build all their racing strategy based on safe data and predictions.

Official races have stringent safety regulations, and AI must adhere to ethical and legal standards. These standards include various requirements related to privacy, security, fairness and more. The regulations encourage safe, fair and responsible use and are intended to build trust in AI systems, like the trust between a racing team and its fans.

While there is yet to be a regulation specifically focused on agentic AI, there are a patchwork of existing and emerging regulations for AI that apply to agents. The ability to audit and track an agent's interactions with data, tools and users can assist with compliance requirements. Race cars have computers that constantly monitor safety and performance and provide postmortems after a race.

If there is a question about whether a team broke the rules, there is data available to review. With most AI, especially agents, tracking and metadata capture are not typically automated. AI governance tools can help with detailed audit trails of everything the agent did to understand what happened, making deploying agents less of a black box. 

Organizations can build their own agents or purchase them from a third-party provider. Once they are comfortable with agentic AI and have the proper processes in place, they can find a multitude of agentic use cases. As organizations increase their usage of agents, tracking all the agents in the organization and mapping can be challenging.

Racing teams have various cars to choose from in their development facility and garage. Their design team selects the right engine, chassis, tires and drivers for each race. However, there can be some crossover in terms of parts, so knowing what tools and components each team has access to across the portfolio is helpful.

Here is where a governed agentic catalog can assist organizations in understanding what is within their “garage” and “toolbox” of agents. Significant time is saved when an agent, agentic tools or workflows can be repurposed or modified from one use case to another. The governed agentic catalog ensures that vetted and trusted tools and agents are used for building agentic AI in each organization. It can also track the utilization of agents across the portfolio to understand which ones are most successful and which are most versatile.

The parallels between agentic AI and racing are more than just an extended metaphor. They can help us better understand and appreciate the complexities and nuances of AI governance, much like a racing enthusiast appreciates the intricacies of a motorsport race.

You can count on IBM to innovate and provide solutions for what matters the most. Today, there is tremendous potential with agentic AI, which is why we’re focused on launching improvements to watsonx.governance® within this area and broadly. You can also find governance within  watsonx Orchestrate® and watsonx.ai ® .

watsonx.governance helps govern, manage and evaluate agentic AI with a series of new features throughout the AI lifecycle. We are also releasing improvements around expediting