In a bustling street, a pair of red and white concrete road barriers stand against a light blue background. These barriers on the highway, much like the ones we're familiar with, serve to protect vehicles from veering off course and into danger. With the emergence of generative AI (gen AI), the concept of guardrails extends to systems designed to ensure that a company's AI tools, especially large language models (LLMs), operate in alignment with organizational standards, policies, and values.

Get to Know and Engage with Senior McKinsey Experts on AI Guardrails

Lareina Yee, a senior partner in McKinsey's Bay Area office, along with Roger Roberts (a partner), Mara Pometti (a consultant in the London office), and Stephen Xu (a senior director of product management in the Toronto office), are here to guide us through the world of AI guardrails.

Benefits of AI Guardrails

Privacy and security are crucial aspects. AI systems are vulnerable to attacks by malicious actors who can manipulate AI-generated outcomes. Guardrails act as a shield, safeguarding organizations and their customers. Regulatory compliance is another key benefit. As government scrutiny of AI increases, guardrails help organizations ensure their AI systems adhere to existing and emerging laws and standards, reducing the risk of legal penalties. Trust is paramount, and guardrails enable continuous monitoring and review of AI-generated outputs, minimizing the risk of errant content being released.

For instance, imagine a healthcare organization using an AI system to diagnose patients. Without proper guardrails, the system might generate inaccurate or misleading diagnoses, putting patients at risk. But with guardrails in place, the system can be monitored and corrected in real-time, ensuring the accuracy and reliability of diagnoses.

Another example is in the e-commerce industry. Guardrails can prevent the sale of counterfeit products by filtering out inappropriate or inaccurate product information generated by AI. This helps build trust with customers and protects the reputation of the business.

Main Types of AI Guardrails

Appropriateness guardrails check for toxic, harmful, biased, or stereotypical content and filter it out before it reaches customers. For example, in a social media platform, these guardrails can prevent the spread of hate speech and offensive content.Hallucination guardrails ensure that AI-generated content is factually correct and not misleading. Say a news organization uses an AI to generate articles; these guardrails would prevent the inclusion of false information.Regulatory-compliance guardrails validate that generated content meets regulatory requirements. In the finance industry, for instance, these guardrails would ensure that financial advice generated by AI complies with relevant regulations.Alignment guardrails ensure that generated content aligns with user expectations and maintains brand consistency. For a brand's customer service chatbot, these guardrails would ensure that the responses are in line with the brand's tone and values.Validation guardrails check if generated content meets specific criteria. If a piece of content fails the validation, it can be funneled into a correction loop. This helps maintain the quality of AI-generated content.

Take a content management system as an example. By implementing these different types of guardrails, the system can ensure that the content published is appropriate, accurate, compliant, and aligned with the organization's goals.

Another instance could be in an educational setting. Guardrails can prevent AI-generated educational materials from containing biases or incorrect information, providing students with high-quality learning resources.

How Guardrails Work

Guardrails are built using various techniques, from rule-based systems to LLMs. Most guardrails are fully deterministic, meaning they produce the same output for the same input. They work by performing a range of tasks such as classification, semantic validation, and detection of personally identifiable information leaks.The checker scans AI-generated content to detect errors and flag issues like offensive language or biased responses. It acts as the first line of defense. Once an issue is identified, the corrector refines, corrects, and improves the output. The rail manages the interaction between the checker and corrector, triggering corrections when needed and logging the processes for analysis. The guard interacts with all components, coordinating and managing the entire process.

For example, in a chatbot application, the checker might detect a misspelled word in the AI's response. The corrector would then correct the spelling, and the rail would ensure that the corrected response is sent back to the user. This iterative process ensures the quality of the chatbot's responses.

In a legal document generation system, the guardrails would ensure that the generated documents are accurate, compliant with legal requirements, and free from biases. This is crucial in ensuring the fairness and integrity of legal processes.

How AI Guardrails Generate Value

AI guardrails not only help meet compliance and ethical requirements but also create a competitive advantage. They help build trust with customers and avoid costly legal issues. By using AI more responsibly, organizations can attract and retain top talent.For instance, a manufacturing company that implements AI guardrails in its production processes can ensure the quality and safety of its products. This builds trust with customers and gives the company a competitive edge in the market.ING, a financial-services company, developed an AI chatbot with guardrails to ensure accurate and safe customer interactions. The guardrails filtered out sensitive information and risky advice, while ensuring compliance with regulatory standards. This not only protected the customers but also enhanced the company's reputation.

Another example is in the logistics industry. AI guardrails can optimize delivery routes and ensure the timely and accurate delivery of goods. This improves customer satisfaction and increases the efficiency of the logistics operations.

How to Deploy AI Guardrails at Scale

Design guardrails with multidisciplinary teams that include legal experts. Define content quality metrics tailored to business goals and regulations. Adopt a modular approach to build reconfigurable components that can be easily embedded and scaled in existing systems. Take a dynamic approach by setting up rule-based guardrails with dynamic baselines that can change based on different variables. Steer with existing regulatory frameworks and develop new capabilities and roles for practitioners accountable for model outcomes.

For example, a large e-commerce company can form a team consisting of engineers, legal experts, and ethicists to design and implement AI guardrails across its platform. By defining specific metrics for content quality, such as product descriptions' accuracy and compliance, the company can ensure the consistency and reliability of its offerings.

In a healthcare system, deploying AI guardrails at scale requires collaboration between IT teams, medical professionals, and compliance officers. This ensures that the AI systems used in healthcare are safe, accurate, and compliant with medical regulations.

The rapid growth of AI has made compliance more complex for companies. Guardrails can help companies manage risks and foster innovation. By incorporating guardrails into various processes like product development, organizations can better handle AI-related crises and create a safer environment for AI-related activities.