AI is no longer just answering questions. It is beginning to make decisions and carry out transactions on our behalf. But as banks and financial institutions give AI more responsibility, who makes sure it stays under control? Joining me here in Dubai is Evin McMullen, Co-Founder of AmanTrue, to discuss how we can make autonomous AI safer, more secure, and accountable. Thank you for joining us.
It is an honour to be here today. Thank you so much for having me.
As AI moves from assisting people to making decisions on their behalf, what changes for banks and financial institutions?
In the advent of AI agents, we can now have autonomous systems that both initiate and execute transactions more than ever before. For financial institutions, this introduces a new security risk and a need for monitoring, control, and management. Banks need to ask three questions: which agent acted, what did it do, and who authorised it? The new considerations for financial institutions need to mirror the management and monitoring of human employees — ensuring accountability and alignment is in place, requiring evidence rather than assurances. As we know from legal precedent and government mandates around the world, accountability does not move from humans to agents. It continues to sit at executive and board level. Those individuals need appropriate transparency and auditability of agent actions.
AmanTrue focuses on controlling AI agents. How is that different from simply verifying the identity of an agent?
Identity answers the question: which one is it? But that is not enough. With AmanTrue, we also assign unique identities to AI agents that are directly linked to and derived from the identity of the humans and businesses that deploy them — and therefore are accountable for their actions. Identity determines which one. Control defines the bounds of what that identified agent is able to do — anticipating tasks, defining boundaries around permissible actions, and ensuring monitoring occurs before an action is taken, not only afterward.
How can a bank ensure that an AI agent only accesses information and performs transactions it is actually authorised to handle?
In the same way that employees within a financial institution are limited in the data they can access and the policies they can enact, financial institutions must apply similar parameters to their agents. Providing more than just monitoring — but actual management systems involving the individuals and leadership who are enabling the deployment of these AI agents — is vital.
Which financial decisions should AI never be allowed to make independently?
Decisions that meaningfully impact the experience and opportunities of human beings should be considered — as the government here in the UAE has described — a red line. Not only should humans be in the loop of this agentic process, but I would argue that humans should be in the lead. We are not yet at a point of development where we have 100% assurance of likely outcomes. Human discretion, oversight, and policy management continues to be essential.
Can zero knowledge proofs help banks verify information without exposing sensitive customer data?
A zero knowledge proof is a method to compute on or learn about data without seeing the data itself. If an agent wanted to verify whether I am over 18, it does not need to see my entire Emirates ID with my birthdate, home address, and personal traits. Instead, the agent can ask the question — is Evin over the age of 18? — and receive a mathematically verifiable answer based on information from a trusted government source. Zero knowledge proofs have already been proven as a useful way to minimise data exposure at a financial institution level and a nation state level. This is not a frontier hypothetical — it is proven at scale in production and ready for enterprise and government adoption here in the region.
The UAE aims to introduce agentic AI across half its federal government services within two years. What is driving that ambition and what are the biggest risks?
More than 70% of working age people in the UAE are already using AI in their work and personal lives — far outpacing many other jurisdictions. The risks are real. When it comes to how AI agents are applied, minimising the extent to which they are exposed to sensitive data also minimises the opportunities for misuse. Mindful setup and design of agents and AI-enabled systems is crucial. The government’s guidance on implementation is a sterling example for other jurisdictions around the world.
Evin, it has been great having you with us.
Always a pleasure. Thank you so much.
