How do you trust AI to move money safely and reliably

Key takeaways from AI-VOLUTION on the topic “How do you trust AI to move money safely and reliably” In this session, Tyllen Bicakcic, CEO & Co-Founder of Payman joined Rocky Yu, Founder & CEO of AGI House) to unpack how Payman is building a “Trust Layer” between AI agents and the financial system.

AI is about “probabilities,” but finance demands “certainty.”

Tyllen points out a critical gap: today’s banking APIs are built in a deterministic way — they require a clear, explicit click or command for an action to be executed — while AI operates in a probabilistic way, where it chooses which tools to use and what actions to take.

The risk: if we let AI make financial decisions without protection, even a tiny mistake can lead to massive losses. That’s why we need strong “guardrails” — safety rails that tightly control what AI can and cannot do with money.

“Trust Layer” as the solution

Instead of forcing AI developers to build all the financial safety mechanisms themselves — which is complex and time-consuming — Payman provides a platform that acts as the intermediary, handling this trust and safety layer. They break the protection into two layers:

Layer 1: Authorization & Permissions

Clearly define:

• Who is actually giving the order
• How much this AI agent is allowed to spend (spend limits)
  Which recipients it’s allowed to pay (approved payees)
• What information the AI is allowed to share (e.g. bank account numbers)
  

Layer 2: Monitoring & Audit

• The system reviews the prompt history to detect prompt injection or malicious instructions hidden in the conversation
• There is an audit log so you can trace back and understand why the AI decided to execute a specific payment or transaction

Enterprise tools (Enterprise controls)
To meet enterprise-level confidence and security, Payman provides key features such as:

• Kill Switch: An emergency button to immediately stop an agent’s operations if something goes wrong
  
• Version Control: Manage and lock agent versions to prevent updates that could introduce bugs or unexpected behavior
  
• Scoped Permissions: Separate read (view-only) permissions from write/transfer permissions, so not every agent that can read data is allowed to move money

Regulatory strategy: not building a new bank, but layering on top

One key lesson Tyllen learned is that giving an AI its own bank account directly (a direct bank account) would likely classify it as a money transmitter, which brings significant legal and regulatory complexity.

So instead, Payman chose to act as a “layer” over existing bank accounts and infrastructure — an overlay on top of current systems.

The future of the “Pagent” (Payment Agent)

Tyllen coined a new term: “Pagent” – an AI agent specialized purely in payments.
In the future, organizations won’t rely on a single AI to do everything. Instead, they’ll use multiple AI agents working together.

When it comes time to actually move money, the task will be handed off to the Pagent, which operates with the highest level of security and control to handle payment execution.

Conclusion

The key to unlocking AI’s full potential in finance isn’t just speed – it’s mitigated risk. Once we have solid systems for authorization, monitoring, and strict policies in place, we can finally transform AI from an experimental tool into a production-grade system that can safely make payments on behalf of humans and generate real economic value.