If you mention “governance” to a business leader, they likely start to think about restrictions and controls – in other words, putting on the brakes. That perspective is incomplete, especially when it comes to agentic AI.
Agentic AI can bring new levels of productivity and performance by assessing situations, gathering and processing data to problem-solve, and autonomously executing tasks with minimal human input. Some businesses have already started using it to reinvent core processes like finance, linking multiple AI agents and models from different vendors. But leaders are also understandably concerned about ensuring these increasingly complex solutions are properly governed, compliant and function as the business needs them to function.
In the rapidly evolving world of AI, governance is no longer merely about applying the metaphorical brakes to avoid obstacles or ensure compliance. Instead, realizing the full value of agentic AI demands a holistic approach – charting a clear course, managing resources effectively and maintaining systems diligently.
In fact, managing agentic AI is much like managing a car.
Identifying the Driver
Just like a car needs a driver, enterprises need a person or group of people who are accountable for their AI and how it travels through the world. The stakes are very high when businesses use agentic AI to support how they engage customers, manage the critical flow of goods in their supply chain, or develop new products. Doing this well requires a funded mandate to ensure that AI solutions are governed, fair and maintained appropriately.
Improving the overall workforce’s AI literacy is also key. Every employee will start to interact with AI agents, not just technologists. Businesses need employees who are trained and empowered to ask critical questions about AI’s outputs and the data that trained it. When multiple AI agents are working together to complete a task, skilled domain experts should oversee their activity, recognize how each link in the chain could fail and help mitigate risks.
Choosing Your Destination
Just as a car needs direction, an effective AI strategy begins with choosing your destination — aligning your AI strategy with your business objectives.
Ask questions like: What key growth objectives could AI help us achieve? What tasks do people do best, and where does AI bring an edge? If I apply AI in this process, how might I reverse the outcome if needed? Is the process well documented with a clear start and end? As you do, keep in mind the rules of the road, like your own company’s principles for AI and the rapidly evolving regulatory landscape.
As part of this, establish an AI Ethics Board that can serve as your GPS, guiding you towards the most beneficial and responsible application of AI. This should include experts from various business domains and functional areas, including legal, HR and compliance, who understand how to work with AI technologies and evaluate their impact. Together, these experts can identify potential risks of AI solutions, determine acceptable risk levels, and ensure they’re meeting requirements like explainable outputs and test-retest reliability.
Managing Your Fuel
A car without proper fuel won’t get far, and the same is true for agentic AI. The “fuel” in this case is data — clean, relevant, representative and accessible to the AI-powered process it’s supporting.
Leaders should focus on data provenance and attribution, quality, accuracy and representativeness; appropriate permissions and consent; and compliance with data protection regulations. This isn’t simple in today’s reality of data silos and complex back-end systems.
Poor data management can lead to biased outputs, privacy violations and ineffective AI systems. By contrast, well-governed data powers agentic AI to deliver reliable, trustworthy results.
Acceleration
Knowing when and how to accelerate is crucial. Before stepping on the gas, identify use cases that align with business goals and maximize value, while mitigating risk. Use cases should be well bounded, with a clear starting point and destination, and well documented, including clarity on what data is being used and if it’s sensitive data. They should also be reversible, supported by a reliable “brake system” that lets you reverse and course-correct when required.
A good example is generating reports from clean, structured data. Imagine an AI agent that monitors a database, analyzes new data, generates a report using a template and then distributes that report in pre-defined internal channels. This use case drives efficiency and productivity but also has strong governance, since the data is sanitized in advance and the report form is known and tested.
Maintenance
Just as cars require regular brake checks or tire rotation, agentic AI systems need ongoing monitoring and care.
If you’re replacing a manual, complex business process that’s essential to how your business functions with a newly re-engineered process that depends on agentic AI, you want to be confident that the process won’t break down. Without proper maintenance, even the most sophisticated AI systems can deteriorate, developing biases or inefficiencies that compromise their value and potentially create risks.
You’ll need end-to-end monitoring for machine learning and generative AI models, checking for model drift and providing alerts when inappropriate language or personally identifiable information comes up. This is especially important with agentic AI systems that operate more autonomously, as unchecked drift can lead to risks like regulatory noncompliance, data privacy breaches or threats to your company brand and reputation.
You also need to monitor and manage cloud costs underlying these solutions. Imagine an AI agent stuck in an infinite loop, repeatedly calling an LLM, because it doesn’t have a guardrail specifying when to pause and consult a human operator.
By viewing AI governance as managing a complete vehicle rather than just the brakes, leaders can take a more effective and balanced approach to scaling agentic AI. Organizations that master all aspects of the “vehicle” will be better positioned to harness the transformative potential of agentic AI while mitigating risks.
