The new world of work isn’t just adding increasingly efficient digital tools anymore, it is effectively adding full digital colleagues. As organisations integrate AI agents into daily operations, the definition of the workforce is changing into more of a machine-person hybrid. According to Anna Collard, SVP content strategy & CISO advisor for KnowBe4 Africa, human risk management must evolve to govern the behaviour of these agents with the same rigour applied to human employees.
When generative AI first burst onto the scene in 2022, chatbots were passive assistants. “We have now moved into the era of agentic AI – members of the organisation increasingly capable of autonomous initiative,” says Anna Collard, SVP of content strategy and CISO advisor at KnowBe4 Africa. Unlike standard large language models (LLMs), these agents identify and solve problems independently. With nearly 40% of enterprise applications expected to have task-specific agents by the end of 2026, organisations are facing a new kind of security pressure.
“Consider the customer support agent that can access internal databases and update records, or the software development agent that writes, tests, and submits code. These agents require human intervention only for unusual cases or final approvals,” explains Collard. “This is why we must stop viewing AI as a tool and start seeing it as a colleague – and like any new recruit, they require a code of conduct.”
Managing behavioural risks and rogue robots
But if AI agents are now colleagues, how much authority should they be given? New recruits to an organisation are typically not given free rein without supervision. “One risk with AI agents is automation bias,” explains Collard. “Employees may stop questioning the accuracy or security of an agent’s output because it is fast and authoritative. This leads to a degradation of human oversight as people delegate their critical thinking to a machine.”
We have already seen the fallout of bots getting it wrong. Air Canada’s chatbot famously hallucinated a refund policy that did not exist, forcing the airline to honour the bot’s “lie” in court. In another instance, an executive using the digital assistant OpenClaw to manage her emails watched as the bot deleted her entire inbox. These are not technical “bugs” in the traditional sense at all, and rather behavioural failures of autonomous agents operating without sufficient guardrails.
The new handbook chapter
Just as a human employee signs a code of conduct, AI agents require a framework for operational boundaries. One of the most critical requirements relates to data-leakage controls. “Because agents learn from the data they process, there must be strict rules to prevent them from accidentally gaining access to unauthorised data or exposing sensitive information, like PII or intellectual property,” Collard emphasises.
These controls must focus on the interaction layer. Employees should be trained to not upload or give access to sensitive data – such as customer ID numbers or proprietary code – to an unvetted agent. “On the technical side, even trusted or internal agents must be given ‘least privilege’ access, ensuring they can only see the data essential for their specific task,” Collard cautions.
Why this matters for HRM
Collard believes human risk management (HRM) should now encompass AI agents as part of the framework of people, technology and systems used to bolster organisations’ cybersecurity posture. “Human risk now inherently means managing how humans interact with, train and oversee their AI counterparts,” she states. “AI-generated output – be it code, emails or financial data – should not be sent without human review for those cases that require full confidence. Similarly, companies should explicitly define which agents have the authority to ‘speak’ for them and what their rights are.”
In this hybrid environment of human and machine, trained humans remain an organisation’s best cybersecurity asset. “Employees are still the best line of defence,” she concludes. “Organisations must move from awareness to behavioural change, reminding employees that they, not their AI agents, are accountable for the final click.”
