EXCLUSIVE: "Banking Beyond The Prompt" - Suresh Ramamurthi, NetXD and Lon Huffman, LJ Huffman Ventures in 'The Paytech Magazine'

Would you trust your AI agent to instruct your bank? Many of your customers already are – and the banks that make it safe first will keep them, says NetXD

Many of us are wary of delegating financial decisions to AI agents, but the picture is changing fast. Customers are already handing them their passwords. Eighty-two per cent of consumers across 13 markets trust AI to compare loans, and more than half would let an agent apply for credit, according to research from Experian and Forrester Consulting, published in July. Another 23 per cent said they would let an agent act under pre-agreed rules and five per cent would give it full autonomy. The appetite is there; the safeguards are another matter.

“Consumers are starting to say, ‘Why can’t my bank do these things for me, if I can get Claude or ChatGPT to do it?’ We’re already seeing individuals give their bank account passwords to these agents, which I’m not so sure is a wise move,” says investor and fintech entrepreneur Lon Huffman. 

Huffman has spent his career developing customer-facing banking technology. The problem, he believes, is less about AI than the plumbing underneath it: core systems, online banking, lending and payments separated by ‘a lot of walls’.

“All that data frankly doesn’t exist in one place at any financial institution,” he says. “The larger institutions are spending tremendous amounts of money today trying to get that data into one place so they can use AI, or even just do analytics.”

Which is why an encrypted, customer-held memory layer that works with any AI agent could be a game changer, Huffman believes. LIAM is just that. A Large Intelligent Agentic Memory, it has been built by NetXD, a banking infrastructure provider whose programmable ledger, payments hub and AI tools are used by banks, credit unions, neobanks and fintechs in the US and 12 other countries across the Middle East, Latin America, Europe and Asia.

LIAM addresses Huffman’s security concerns from the start, and that has a lot to do with what NetXD Chairman and Founder Suresh Ramamurthi learned from building banking infrastructure at scale: trust, compliance, risk and reconciliation must be locked down ‘before the money moves’. A large global bank’s security team told Ramamurthi that banks fix high-priority vulnerabilities first and let medium and low ones wait. 

“From an AI perspective, everything is a vulnerability,” he says. Huffman puts it this way: “We’re really good at locking our doors and closing the windows on the first floor. AI is really good at getting to the bathroom window on the second floor.”

But NetXD was built with zero trust for machine commerce before AI. That’s handy when AI is starting to make its way into other people’s systems without permission. And it targeted the hardest use cases first: institutional clients ‘banking the customers of customers of customers’. The platform now supports institutional clients that need a million or more virtual accounts, and NetXD has processed more than $100billion in payments, ranking it among the top five in US real-time payments.

Once the use case in institutional banking was proven, business and retail followed, and when generative AI arrived, Ramamurthi considered building a model.

 “We were very tempted to start post-training or pre-training LLMs, and, given my background, it wasn’t hard to figure out how to do that,” he says. “Then we realised that LLMs are probabilistic, and banking is a deterministic business. So we needed to  build long-term memory that was deterministic.” But still, the dual issues of security and liability were paramount.

“I’m also chairman of a bank, and this is the biggest question in my mind,” says Ramamurthi. “I didn’t want to wake up in the middle of the night worrying about it. So we addressed it in the design.”

LIAM holds personal account data and preferences, but it belongs to the customer, not the agent or bank, says Ramamurthi. In a real-world scenario, the bank continues to do what it has always done: give customers their own data. It’s just the vehicle that has changed.

“Thirty years ago, I gave your data to you in a passbook. Ten years ago I gave it to you as a CSV or PDF file in online banking. Today I’m giving it to you via an API,” says Ramamurthi. “Anything that came into the bank had to have the customer’s signature on it, as it does today. So the bank doesn’t have to change anything.”

The customer connects their data to an agent of their choice, with LIAM supplying the context. Whatever passes between customer and agent stays there.

“The bank has no awareness, just as today it doesn’t know when you upload your CSV file into a frontier model,” Ramamurthi adds.

And the agent never instructs the bank. It sends its suggestion to the customer, where it appears in mobile or online banking as an action they can take if they choose. Nothing reaches the bank until they approve it, and no transaction can be processed unless it is signed with the customer’s private key.

“The bank only honours [the customer’s] signature. It doesn’t honour the AI’s request,” says Ramamurthi. “The agents never have direct access to the bank’s APIs.”

Customers can revoke access, every action leaves an immutable audit trail, and real-time risk and fraud decisioning and kill switches are built into the platform. NetXD says it is the first platform offering zero-trust security on this banking data pipe.

Huffman’s view is that this is the only order in which agentic banking can work. 

“At the point at which the transaction actually happens, you need the permissions and the audit trail,” he says. “It can’t happen afterwards, where you check at night to see what’s gone on. You can’t have the consumer’s desire to do it without the trust and permissions set up appropriately on the back end.”

Data from every account a customer holds can be mirrored for them through the platform, and the memory stays with them if they change agents. They could even run two agents side by side, Ramamurthi suggests, and ‘make them fight it out’ over which gives better advice before choosing whether to act. For banks, that is uncomfortable. If agents recommend moving money elsewhere, the bank risks becoming ‘an execution utility for your transactions’.

“We’ve noticed that not all legacy banks want you to have an easy way to move money out, or bring money in,” he says. “That is going to change in the industry, and the first movers have an advantage today.”

Huffman says that in his experience, banks adopt customer-facing tools as ‘a defensive play’. They want to maintain the relationship with their customers rather than lose it to fintechs.

But he also argues that they need not start with the most ambitious applications. There is an opportunity to demonstrate the value of agentic banking without handing AI responsibility for complex financial decisions. Huffman suggests starting ‘inside the four walls’ of a bank, with relatively low-risk moves, such as sweeping spare savings into a higher-interest account.

And the advantage of NetXD’s agentic layer is its flexibility. Banks can connect it to existing APIs, run its programmable ledger alongside their core banking system, or launch a separate digital bank using the full NetXD stack. Each approach introduces new capabilities without necessarily replacing existing infrastructure.

One US bank is already implementing the platform, says NetXD. Ramamurthi rests his case for embracing agentic banking on what banks will have left once payments are commoditised.

“In the next few years, maybe banks offer APIs or MCPs [Model Context Protocol servers] and agentic access to data that lets the customer move quickly. Maybe that creates a switching cost, like it did 40 years ago when banks offered VPNs for large enterprise customers to connect into the banking system.

“But what we’ve built into the substrate is programmability. So it’s not a dumb utility, because enterprise customers and consumers alike need programmability. And the bank will plug into the customer’s agent with the highest form of quantum-safe security.”

Ultimately, it could transform CX and elevate every account to a private banking experience.

“What does a private banking customer get?” says Ramamurthi. “A human being who gives them options to make more money with their money and safeguard it. Those are just programmable scripts today.

“Henry Ford made sure the car was affordable to everyone. I believe a successful bank five years from now will make private banking available to everybody. Meeting a human will be a quaint experience. The reality is your agent is going to connect to your bank, and every account will be run like a private bank’s.”