Dan McCalla - Copywriter
Freelance writer in the UK, covering B2B marketing copy, editorial and just about anything else...
Blue and Black Modern Cyber Security Presentation.jpg

Security Scribblings

Why Humans Like You Are Vital For Agentic AI Controls

It’s clear by now that responsible business use of AI agents needs some form of control.

And that’s one area where humans are probably better-placed than AI itself.

To understand why humans are so important for keeping agentic AI use in check, it’s important to distinguish exactly what those controls are.

For example, an everyday user workflow may involve a request for AI use, which can then be verified and authenticated if it’s appropriate within the context of the business. Or it could be ensuring that the only AI operations that can take place within business systems are those specifically authorised - and nothing beyond that. 

Of course, there’s no technical reason why a different AI platform couldn’t take care of those checks and balances itself, once it’s trained with the right contextual business information and so on. But in a climate where just one access or workflow error could cause untold damage from a security perspective, can you trust AI to get it right 100% of the time?

  • As business needs change, AI agent use patterns change - and AI might not understand those nuanced shifts right away

  • Recent well-publicised cases of AI development taking its own action means there is a risk (even a small one) of AI deciding its own controls

  • Many AI models have their own safeguards built-in, but that doesn’t provide an extra degree of control at the point of the data involved

So this is why humans have an essential role to play - most likely in conjunction with AI - in preventing any unauthorised or dangerous use of agentic AI within a business. Designated owners of AI controls should be able to:

  • Ensure every AI agent is running on the principle of least privilege

  • Enforce approval processes for sensitive actions, such as those involving personal data or financial transactions

  • Regularly log, monitor and review agent activity to keep them aligned with appropriate business use

  • Establish a rapid-response shutdown process to terminate agents that ‘go off the rails’ before they can cause any major business impact

Are there any other ways you think humans can add control and balance to business use of agentic AI? Join the conversation on LinkedIn and let me know. 

Dan McCalla