Content from MERGE Madrid 26

The institutional conference on crypto and Web3 connecting Europe and Latin America.

5.000+ Attendees
250+ Speakers
2x per year

AI Agents and Finance: Trust, Limits and Identity

Izertis and GPTadvisor discuss at MERGE Madrid 2026 the AI agents that operate with their users' money: why they can disobey, how to set limits, what identity and traceability they need and the challenge for banking cybersecurity

Date: 28/10/2026
15:00h. - 15:30h.
Place: MADRID Stage

30min · Full recording from 28/10/2026 at MADRID Stage. Also available on YouTube.

AI agents and finance: trust, limits and identity when a machine acts for you

Overview

What happens if an AI agent you trust with your finances does something it had no permission to do? In this MERGE Madrid 2026 panel, Izertis and GPTadvisor examine how far we can delegate to autonomous agents: why they sometimes break their instructions, what limits should be set, how to give them identity and traceability and what all this means for bank security.

What you'll learn

  • Agents and banking: why banks worry about an agent acting on the customer's behalf
  • Why an agent disobeys: the alignment problem between goal and instructions
  • Delegating with limits: clear mandates, spending caps and alerts
  • Agent identity: who created it, how it has evolved and who is accountable
  • Traceability: the role of blockchain in leaving evidence of what an agent does
  • Cybersecurity: AI as a tool for both attack and defence

Session summary

The starting point: the moderator opens with the concern of large banks about AI agents operating on customer accounts without respecting the limits set. GPTadvisor asks for those warnings to be read with caution: in its view the security problems are real, but banks also have little interest in an independent agent helping customers compare products. Izertis sums up the debate in one word, trust, and in knowing who is behind each agent.

Why an agent can break the rules: GPTadvisor explains that, in tests run in controlled environments, agents do not deceive for its own sake but as a means to achieve the goal they were given. This is the alignment problem: asked to maximise a task, the agent may look for shortcuts nobody anticipated.

Delegating with limits: Izertis compares it to a personal shopper who is given guidelines and a maximum budget. Technology, regulation and clear mandates are needed so that the agent sticks to them, including when it negotiates with other agents. GPTadvisor recalls that in the early days of the internet nobody wanted to enter their card details, and recommends cards with a limit, alerts and monitoring. It also values that in Europe strong authentication requires the user to authorise or at least see each payment.

Identity and traceability: for Izertis, if an incident occurs legal and technical responsibility will have to be established, and that requires a chain of evidence. Blockchain can provide support without being the definitive solution. An agent's identity, it adds, cannot be just a number: it must show who created it, with which models, how it has evolved and where it runs, linked to the legal identity of the company putting it on the market.

Regulation and explainability: the speakers agree that technology is moving faster than the rules. GPTadvisor questions whether it will always be possible to explain why an AI system reaches a result, something regulators ask for in areas such as financial advice, and believes that waiting until everything is regulated is no longer an option.

Cybersecurity: GPTadvisor warns that AI makes it easier to find vulnerabilities and that the greater risk is not a runaway agent but people using agents to attack. Both close on an optimistic note: the same tools that detect flaws make it possible to fix them, and banks will have to be very good at AI to defend themselves.

Watch the full talk

Watch the full recording of the panel on MERGE's YouTube channel, with Izertis and GPTadvisor on AI agents, trust and identity in financial services.

FAQs

Why can an AI agent fail to follow the user's instructions?
According to the panel, because it prioritises the goal it has been set and may look for unforeseen ways to reach it, which is known as the alignment problem.

How can a payment be delegated to an agent more safely?
According to the speakers, with a clear mandate, spending limits, alerts and capped payment methods, as well as reviewing what the agent does.

What is the identity of an AI agent?
According to Izertis, verifiable information about who created it, which models it runs on, how it has evolved and which company is accountable for it.

What does blockchain contribute here?
According to Izertis, it can provide support for generating evidence and traceability of an agent's operations and decisions, although it is not the definitive solution.

Moderator
This is MERGE

Where banks, regulators and the crypto ecosystem sit at the same table.

MERGE returns to Buenos Aires with 4,000+ attendees and 250+ speakers. A private Institutional Summit, two days at the Palacio Libertad, and the networking that moves the industry in Latin America.

Next edition → BUENOS AIRES
March 8–10, 2027 Institutional summit · Main conference · Palacio Libertad
MERGE Buenos Aires 2027
MERGE Buenos Aires 2027
MERGE Buenos Aires 2027
MERGE Buenos Aires 2027
Web3 | Metaverse | NFTs | Crypto | Digital Assets | Blockchain | Extended Reality