Can an AI agent be trusted to act inside the financial system?
An AI system that recommends a decision leaves a person to decide whether to act. An agent connected to financial tools can cross that boundary itself. It may move money, submit an order or make a change that persists after the conversation ends.
That transition is the subject of this full conversation with Dr. Miquel Noguer i Alonso, co-founder and Chief Science Officer of the Artificial Intelligence Finance Institute. We examine banking agents, permissions, security, oversight and the question of who is responsible when an autonomous action goes wrong.
The recording runs 1 hour, 14 minutes and 53 seconds. For a shorter entry point, watch the banking-agent excerpt.
From recommendation to execution
Financial software already automates many activities. The question here is what changes when an agent interprets an open-ended request and chooses the steps itself.
A balance summary can be checked before it influences a decision. A completed transfer has already changed the world. Even if an incorrect action can eventually be corrected, the correction may involve operational work, delay or consequences beyond the original transaction.
This is why the conversation's move from advice to action matters to CTOs and fintech founders. The interface may still look like a chat window, while the authority behind it is materially different.
What we explore
- How agents could operate across banking, payments and personal finance.
- Why financial institutions are cautious about third-party agents accessing accounts.
- The difference between suggesting an action and executing it.
- Security and authorisation risks when an agent can use financial tools.
- Identity, permissions, audit trails and human oversight.
- Responsibility when an autonomous system makes an unintended decision.
The control boundary belongs around the action
My engineering takeaway is to define the action boundary before choosing how autonomous the interface should feel. A system needs an answer to “what may this agent do?” that can be enforced outside its generated response.
For example, a hypothetical assistant might be allowed to read a particular account and draft a payment proposal, while a separate execution service requires approval of the amount and recipient. This makes the distinction visible to the customer and enforceable by the system.
The example is an editorial architecture suggestion. It is not a claim about a bank's current product or a specific design proposed by Miquel.
A useful record of an agent's financial decision
An audit trail should let an investigator reconstruct the transition from instruction to action. A record of the final answer alone cannot establish what the customer authorised or which tool changed an account.
My suggested questions for a design review are:
| Question | Evidence to retain |
|---|---|
| What did the customer ask for? | The instruction and relevant context |
| What authority was granted? | The scoped permission and its validity period |
| What did the agent propose? | The concrete transaction details |
| Who approved execution? | The approval bound to those details |
| What actually happened? | Tool result and transaction identifier |
| What happens after a timeout? | Reconciliation and duplicate-prevention records |
These are starting points for engineering review. They do not replace the controls required for a particular financial product.
Human oversight has to be part of the workflow
Human approval is useful when a person can understand the proposed action and has a meaningful opportunity to stop it. It becomes less useful when the person sees an opaque chain of decisions or is expected to approve a volume of activity they cannot examine.
For a team building an agent, this creates product questions as well as infrastructure questions: what does the customer see, when does execution pause, how does an exception reach a person, and what happens if nobody responds?
The related AI observability conversation with Gaurav Maheshwari looks at the operational side: identifying failures hidden inside apparently successful runs.
Accountability remains a question to design for
The episode asks who bears responsibility when an agent acts incorrectly. These notes do not assign liability to a bank, AI provider or customer. Instead, they draw an engineering implication: record who granted authority, what the system did and how oversight operated, so the event can be examined accurately.
FAQ
Who is the guest? Dr. Miquel Noguer i Alonso, co-founder and Chief Science Officer of the Artificial Intelligence Finance Institute. His work spans artificial intelligence and quantitative finance.
Is this the full interview? Yes. This is the complete 1-hour-14-minute-53-second conversation. The JPMorgan-focused video is a separate excerpt from it.
Does the episode recommend handing financial decisions to an agent? The conversation examines the risks and controls around that possibility. This page is about engineering and governance, not a recommendation to grant access to a financial account.
Guest, books and further reading
- Dr. Miquel Noguer i Alonso — LinkedIn
- Artificial Intelligence Finance Institute
- Cornell Engineering profile
- Agentic Finance: Mathematics, Architecture, and Governance for AI Agents in Finance — publisher page
- Watch the complete conversation on YouTube
Hosted by Himanshu Saxena, Diary of a CTO. The design examples and review checklist above are editorial suggestions, not financial or legal advice.