top of page

Managing the Risks of Disruptive AI: Treat it like a new starter

5 hours ago
7 min read

AI the new employee

By David Fox

Director of Fox Risk Management

19 November 2026

 

The discussion


Recently, I had the privilege of sitting in a room with thought leaders and experts from academia, insurance, technology, governance and risk management for an important and insightful discussion of risk management in the age of AI. The conversation left me with one key thought: the uncertainty around AI is real, but it does not stop us from recognising its risks, and we can manage them by treating AI much as we treat a new employee.


As can be expected from a diverse group of experts, views and opinions differed slightly, depending on the lens through which the topic was viewed. What everyone did agree on is that AI is here to stay, and that we all need to become familiar with its benefits and its shortcomings, not just for businesses but for human society as well.


Some people raised the concern that this was unprecedented uncertainty, and that we could not identify the risks to our society or our business. This is the one point on which I hold a different view. I will first look at how earlier technologies, including social media, played out, then set out the risks AI already presents, and finally propose a practical way for executives and Boards to manage them.


Lessons from earlier technologies


The term AI can be used very broadly or for a narrow band of technology. In a discussion yesterday with my 28-year-old son, he described AI very broadly, as covering almost all digital algorithmic processes, including the sort driven by if/then code statements. He is a bright, well-educated young man and an avid user of AI. I might almost say he is the perfect consumer for AI, as he likes to refer to ChatGPT as if it were his new life guru. While this is an interesting story, it is also a warning. If my son is a typical example of unquestioning adoption by young adults, in the same way as social media was embraced by a young generation, then we are in danger of repeating the same flawed social experiment, one that has caused documented harm to a whole generation, harm that we are still trying to understand and correct. Governments are still struggling to put guardrails in place to reduce the ongoing, unexpected harm social media has caused. Technology is always a double-edged sword. Social media has given our society the instant connection we never had before: families spread across the globe can stay in touch like never before, and news and the truth can be recorded and shared globally in real time. But so too can lies and propaganda.


This is where my view differs. AI is a technological advancement for humanity, just like the

personal computer, the internet and social media. We have seen the impacts these advancements have had on our society, our businesses and our personal lives. Many have been hugely beneficial, but every advancement has also brought harm. Those previous harms are precedents we can use to anticipate the risks AI may pose to ourselves, our businesses and our society. The uncertainty is therefore not unprecedented in kind, and it does not prevent us from recognising the risks.


The risks AI presents


The evidence of harm is already emerging, which is why we need to build guardrails early rather than as a reactive regulatory exercise after the harm has become systemic. Empirical studies link heavy AI use with weaker critical thinking, mediated by cognitive offloading (Gerlich, 2025), and early, preliminary EEG research suggests reduced brain connectivity among frequent AI users (Kosmyna et al., 2025, preprint). There is also the compounding effect of AI on the misinformation already flowing through our social media feeds: we can no longer believe what we read, and we also cannot believe what we see and hear.


AI also amplifies bias. Machine learning models trained on historical data often absorb and reproduce human prejudices, leading to unfair or discriminatory outcomes in high-stakes areas like hiring, credit scoring and legal systems.

For wholehearted adopters like my son, the most harmful impact may be its accuracy and reliability, commonly called hallucinations. AI systems frequently generate confident yet false or misleading information. We humans are hardwired to treat a person's certainty as a shortcut for accuracy, a bias known as the Confidence Heuristic. This is why we fall for fake but confident social media articles, the charming and confident used-car salesperson, and confidently presented AI information.


We tend to focus on the output risks of AI, but there are also legal and financial risks. If an employee makes a decision which results in a legal breach or an adverse customer impact, it is simple to identify who is accountable. But who is accountable if an AI agent takes the same decision and the result is a liability? Vendor contracts tend to be very clear about their lack of liability for algorithmic failures or data loss, which can leave the business carrying the full risk.


Another key risk, seen with every technological advance, is the weaponisation of the very same technological benefit. In September 2025, Anthropic reported that a Chinese state-sponsored group had manipulated Claude Code to attempt to infiltrate roughly 30 global organisations, succeeding in a small number of cases (Anthropic, 2025), something no existing security framework was prepared for. We are also seeing a growing “Dark AI” ecosystem, in which malicious language models are sold on underground forums as subscription services, lowering the skill needed to run phishing and fraud campaigns (Unit 42, 2025).


Taken together, these are not unknowable risks. Loss of critical thinking, amplified misinformation, embedded bias, confident but false output, unclear accountability and the weaponisation of the technology are all variations on harms we have seen with earlier technologies, and each can be named, owned and mitigated. While some risks do not yet have obvious mitigations, that does not mean we cannot deploy practical strategies now to start reducing the threat.


Managing AI like a new starter


Which brings me back to my point: we cannot say that the impact of AI is so uncertain that we are unable to recognise its risks or attempt proactive mitigation. The question is not whether we can see the risks, but how we organise ourselves to manage them.


I believe the best and safest way to approach AI use in our lives and our business is not to think of it as a software system or IT tool, but to treat AI, especially agentic AI, as we would a gifted intern or graduate: someone totally goal-oriented, who does not understand the social norms or the business operating model we take for granted, and who, without enough guidance and limitations, will often go off in unexpected directions to achieve the goal put before them. So, just like a new starter, AI needs an induction into our business culture, to be taught what is acceptable and what is not. It needs a security pass that allows access only to the areas it is authorised to enter, and a mentor to support its integration into the business. It needs a line manager who is responsible for the scope and output of its work and who approves the decisions it makes. And, like any other employee, it needs a delegated level of authority around the type and value of the decisions it can make, and an escalation path when a decision falls outside that authority.


What executives and Boards should do


For executives and Boards, it is important to realise that accountability for the outcomes and deliverables of AI within their business is ultimately theirs. This answers the question of who is accountable when an AI agent makes a poor decision: just like the actions of any other employee, AI can affect the reputation and regulatory obligations of the business, and responsibility cannot be passed to a vendor.


It may be appropriate for a board to create a Chief AI Officer to ensure that the responsibilities and accountabilities associated with the use of AI in the business are not lost in the gaps between departments and systems using AI. The Chief Information Officer, who traditionally focuses on internal IT infrastructure and enterprise applications, is not the right person to be responsible for AI. AI is not like other IT systems, which execute predefined workflows. AI systems analyse complex, high-volume data and autonomously plan multistep goals, so they are much more like an employee than a machine. They need continuous oversight and constant tracking of their sourcing of raw data to ensure they are not drawing on and learning from obsolete or incorrect data, or changing the raw data as they generate outputs that are then absorbed back into their learning. The uncertainty around AI is real, but it is manageable if we govern it with the same discipline we apply to the people we trust with our business.


My message to executives and Boards is to act now rather than wait for certainty. Name a single accountable owner for AI, whether that is a Chief AI Officer or an equivalent hybrid role. Treat every AI system as you would a new starter: give it an induction, limited access, a line manager, a delegated level of authority and an escalation path. Ask your vendors plainly who carries the liability when an algorithm fails. The risks are recognisable, and the guardrails are far easier to build early than after the harm has become systemic.



References

Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv:2506.08872 [preprint]. MIT Media Lab.

Unit 42. (2025, 25 November). The dual-use dilemma of AI: Malicious LLMs. Palo Alto Networks.

 
 
 

Comments


NZ Background.jpg

Ready to strengthen your risk strategy?

bottom of page