This is the next instalment of a multi-part² series highlighting key takeaways from a World Arbitration Update panel on International Arbitration & Emergent Technologies, which took place in Singapore on 22 May 2025.

--

Robert Houston posed a tough question to Yasmin Hussain and myself about (and I'm paraphrasing here!) the rise in the use of agentic AI,³ the possibility of companies facing liability for unexpected acts of AI agents, and what it means for legal practice.

So for example, suppose two companies each use agentic AIs to deal with each other. Then something goes wrong. What next?

I suggest, however, that this is not a particularly novel problem, and we already have potential solutions that are more nuanced than a dogmatic "don't do it".

After all, organisations already have to deal with the consequences of agents acting in unexpected ways.

If what we're concerned about is the prospect of agentic AIs entering into contracts or executing orders, and getting it wrong...

...how is this any different from the current state of play, where organisations rely on human actors, who - let's face it - regularly make mistakes?

I mean, it doesn't make sense for even a mid-sized organisation to require its CFO to sign off on each and every contract, no matter its nature or how big or small.⁴ Which means such authority has to be delegated down the pyramid. And which also means that there is an inherent risk of that authority being exercised in a manner that upper management would not have.

And we already have solutions designed to address this concern, such as:
- internal training;
- authority limits;
- exclusion and limitation of liability clauses;
- errors and omissions insurance;
- legal mechanisms, such as the law of mistake or third-party claims against agents;
and the list goes on.

Now, I'm not suggesting that each and every one of these mechanisms are fit for purpose when it comes to managing the risks of agentic AI.⁵

But I do suggest that we can, by analogy, consider which of these (or other) solutions can be retooled towards agentic AI.

After all, as Simon Chesterman puts it, think of an LLM as "a very smart intern with a drinking problem".

And if this analogy holds with agentic AI, I suppose we'd have to ask ourselves what guardrails we'd put in place to get the most contribution out of such an intern, while managing risk.

--

In the next part, I'll wrap up with some observations on the interplay between AI and arbitration.