GLG Frontier
>
Research

Before an AI Agent Moves Money, It Has to Prove Four Things

Agents in banking are crossing from advice to action. Three practitioners – a security architect who built agentic AI risk frameworks in financial services, an engineer who scaled AI compute at a major chipmaker, and a contact center technology chief – agree on where the real controls belong. They split on how much time banks have to build them.
Banking
Frontier Briefing
GLG Frontier
November 2026
CONTRIBUTORS
GLG Frontier Research Team

Soon, agents in finance won’t just be giving advice; they will be acting on their own. We spoke to three practitioners – a security architect who built agentic AI risk frameworks in financial services, an engineer who scaled AI compute at a major chipmaker, and a contact center technology chief – to understand where controls belong and how much time banks have to build them.

When AI moves from advice to action, risk completely transforms. A wrong answer becomes a lost payment. And because agents act at machine speed, one mistake can repeat across many bills and accounts before anyone notices.

The practitioners agree on the general solution: agents can handle money safely, but they need to operate within boundaries beyond the model’s control. They offered different perspectives on timing and urgency, though. While consumer agents are still some way from routinely moving money, attackers using AI are not, and in the security architect’s view, that gap should set the pace for security.

A Valid Login Proves Less Than You Think

With agents in play, banking security needs to account for more than logins. A hijacked agent can ride a perfectly genuine session.

Before money moves, the experts agreed, a bank should authenticate four separate things:

  • The human who assigned the agent.
  • The agent itself.
  • The permission – the account, action, amount, recipient, and expiry the customer approved.
  • The intent behind this specific transaction.

The former chipmaker engineer described a similar chain from the merchant side Agents act through short-lived, task-scoped tokens, never the customer’s password. In the engineer’s framing, agent identity will look less like a chatbot and more like delegated authorization: a move away from screen scraping toward an OAuth-style grant with payments and logs attached.

That also means an agent has to declare itself. An agent that poses as a browser or a person can’t be scoped, audited, or held to merchant rules.

Draw a Tight Authority Envelope

Not every action carries the same risk, and all three practitioners tier it in similar ways.

The chipmaker engineer calls them “confirmation tiers:” automatic for a routine purchase like paper towels, a confirmation step above that, and required human approval for anything with legal weight, such as changing an insurance policy. The contact center executive draws the same line between bookings, which agents already handle well, and disputes and cancellations, which still need human oversight. His shorthand is that “the pilot stays in the plane.”

Two older banking rules can be useful. The agent that creates a high-risk request should never approve it. And the right to discuss an offer doesn’t equal the right to sign it: an agent can shop for a better insurance rate, sharing only the personal information it needs, but the customer approves the deal.

A simple solution would be a limited agent “wallet:” a dedicated account with a capped balance the agent can’t raise on its own. “Give the agent a controlled spending pocket, not the keys to the entire family saving,” the security architect said.

Keep the Rules Outside the Model

Better prompting alone won’t protect against prompt injection attacks.

The fix is architectural. The chipmaker engineer frames it as accounting for agents: every request and approval traces back to whoever made it. Treat every outside document as untrusted data, never as new authority. Check payment details against the customer’s original instruction and verified recipient records. Run a payment policy service outside the model that blocks forbidden changes before funds move. Keep the original invoice, the proposed payment, and the approved transaction for forensics.

“An invoice may provide facts, but it must never gain the authority to change who gets paid,” the security architect said.

How Close Are Agents, and How Fast Must Controls Be Ready?

The chipmaker engineer argues the building blocks for safe agent commerce – scoped authorization, tokenized payments, and audit logs – largely exist already. In his view, the variable to watch is whether consumers and enterprises hand agents real, repeated workflows.

The contact center executive is more cautious about the consumer side. In his read, agent-placed calls still make up a small share of inbound traffic because consumer agents remain difficult to set up, so business-to-business use comes first. He also expects AI-to-AI phone calls to be a short phase. Once a customer has authenticated an agent with a company, the exchange becomes the equivalent of an API call, and voice drops out. Meanwhile, enterprise upgrades take months; the models calling in change monthly.

The security architect expressed more urgency. Attackers already use AI to write convincing scams, sort stolen data, and find known weaknesses faster than before, and he expects such cases to multiply soon. That’s why one precondition is non-negotiable in his view: no autonomous payments until a bank has fixed critical vulnerabilities and removed easy attack paths.

Who Pays When an Agent Gets It Wrong

Liability remains an open question. The security architect starts with the customer agreement. Coverage depends on the country, the payment type, and whether a transfer counts as authorized. A bank can’t assume an agent’s action equals the customer’s informed consent, so the agreement should define consent, accountability, evidence, and dispute handling before agents move money.

The chipmaker engineer frames it as a three-way question among the user, the agent platform, and the merchant, with no settled answer yet on who arbitrates fraud disputes. He expects underwriting to change, as with self-driving cars. The contact center executive predicts that AI vendors will end up carrying most of the accountability, partly because people hold an AI’s mistakes to a higher standard than a human agent’s.

Whoever pays, the experts agree responsibility should consider a record of the customer’s instruction, the agent’s behavior, and the controls applied at each step.

Our Take: What This Means for Teams Building Agents

For model builders and deployers, the authority layer around reasoning represents a greater challenge than the reasoning itself. A few questions worth testing before an agent touches money:

  • Can the agent act beyond its scoped permission if a document tells it to?
  • Is there any path for the agent to raise its own limits or approve its own requests?
  • Does the agent identify itself as an agent on every channel, including voice, and fail cleanly when a counterparty challenges it?
  • Do high-risk actions route to a separate human channel by default?
  • Can a clear violation be denied and contained without waiting for an analyst?
  • Is there a complete evidence trail from customer instruction to execution?

Timing is key. Evals here need refreshing that matches month-to-month the cadence that detection and evasion leap-frog each other.

These are domain questions as much as technical ones. The right thresholds, edge cases, and dispute rules come from people who have run security, payments, and customer contact inside regulated institutions. That practitioner judgment is what turns a generic agent benchmark into one a bank’s risk team trust.

Go Deeper

This briefing draws on three GLG Library interviews: “AI Agents in Banking and Personal Finance – Sector Blueprint” (October 14, 2026), “Personal AI Agents – Racing Adoption, Resetting Hardware Demand” (October 1, 2026), and “Contact Center AI – When the Caller Is an AI Agent” (October 15, 2026). GLG clients can read the full transcripts in myGLG. To pressure-test your own agent design, GLG can connect you with practitioners in payments security, fraud, identity, and contact center operations through a one-on-one expert call. Teams building evaluations for agentic finance can work with GLG Frontier to design benchmarks and red-team scenarios grounded in real practitioner judgment.

Get connected to the world's leading platform for trusted human expertise

Enter your information below to get in touch with our team.