Pretxt, a company with locations in Aspen, CO, and Dublin, Ireland, has announced the launch of its AI supervision tool, backed by a $5 million seed investment. The funding, sourced from Deciens Capital, Howdy Partners, and Service Provider Capital, positions Pretxt within the growing niche of firms developing technology to prevent AI agents from exceeding their operational boundaries and causing potential harm.
The Rise of AI Supervision
AI systems are distinct from traditional computer programs due to their ability to interpret and respond to complex instructions, often leading to unintended outcomes. This capability of AI to misinterpret instructions can result in agents acting beyond their intended purpose, posing significant risks. To mitigate such risks, limiting the operational scope of AI agents becomes necessary, though overly stringent restrictions could stifle their potential benefits to businesses.
Without innovative supervision solutions, the challenge remains to harness the advantages of AI while preventing rogue behaviors. This necessity drives the development of cybersecurity products that aim to identify and manage potentially harmful actions by AI agents in real-time, ensuring they operate within safe parameters.
Pretxt’s Innovative Approach
Pretxt introduces its AI Supervision tool, designed to oversee AI agents as they function, guiding them back to the intended path before their actions lead to negative business impacts. Aaron Costello, Pretxt’s co-founder and CTO, elaborates that their approach does not diminish an AI agent’s capabilities. Instead, similar to an experienced supervisor overseeing a trainee, Pretxt continuously monitors AI interactions with systems and people, ensuring compliance with business standards.
Costello explains that if an AI agent deviates from its authorized tasks, Pretxt’s system ‘steers’ the agent back on course rather than deactivating it, maintaining operational continuity while ensuring adherence to intended purposes.
Ensuring AI Compliance
The ‘steering’ mechanism employed by Pretxt is context-driven. Upon onboarding, clients connect their AI agents to Pretxt’s tool, allowing the system to learn the agent’s instructions, capabilities, and operational context. Clients specify constraints, such as prohibiting specific actions or data access, which are integrated into Pretxt’s AI to establish a framework for acceptable behavior.
For instance, if a customer service AI agent at a medical practice is instructed not to provide medical advice, Pretxt ensures compliance by monitoring interactions in real-time. If the AI attempts to offer medical advice, Pretxt intervenes by redirecting the agent’s focus without disrupting its primary function, thus preserving service continuity while preventing unauthorized actions.
Pretxt’s AI supervision application is crafted to maintain the operational safety of AI agents, preventing them from breaching predefined constraints while supporting their beneficial roles within organizations.
