Aurélien Bonnel, CTO and Founder of N3XT – Interview Series

Aurélien Bonnel, Chief Technical Officer and Founder of N3XT, is a financial technology executive and engineer with more than 14 years of experience building secure, real-time banking, payments, and capital markets infrastructure. Before founding N3XT, he held senior engineering roles at Deutsche Bank, Nimbla, Symbiont, and SADA, and helped develop the blockchain payments infrastructure used by Signature Bank’s Signet platform. His career has focused on modernizing financial systems through cloud-native architecture, blockchain, automation, and scalable payment technology.

N3XT is a fully reserved bank built around modern infrastructure for instant, programmable business-to-business payments in U.S. dollars. Unlike conventional banks, N3XT does not lend customer deposits, with deposits backed by cash or short-term U.S. Treasuries. The company is also expanding into the intersection of banking and AI through N3XT MCP, a Model Context Protocol based system designed to connect AI agents and assistants with live banking data while maintaining existing permissions and compliance controls. This could enable AI-driven workflows for payment preparation, reporting, financial monitoring, and other corporate finance tasks.

Your career has taken you from pricing and high-frequency trading technology at Deutsche Bank to building blockchain payment infrastructure, leading engineering at Symbiont and ultimately co-founding N3XT. Which problems repeatedly surfaced throughout those roles and convinced you that a new type of bank needed to be built from the ground up?

No matter my personal history, anyone can see the problems that led to the creation of N3XT. We all know the problem of payments being made on a Friday, only to land in the recipient’s bank account on the following Tuesday. The “Available Balance” in that account is different from the balance shown because the payments “received” have not settled and are not usable.

I wondered why. Who benefits from all those delays? It turns out the banks do. Every day of delay is interest they make off the backs of their clients. Banks have had access to the same technology N3XT uses for years, but they’ve not adopted it — perhaps because it would expose their business model.

All of this made it evident to me that the only way to fix this system is to come up with a radically new model, built from the ground up, with new ideas: A narrow foundation, where a bank does not lend and settles payments instantly.

N3XT MCP is designed to connect AI assistants directly with live banking data and workflows. What does the Model Context Protocol make possible that could not be achieved as effectively through conventional banking APIs, enterprise integrations or robotic process automation?

First, it’s worth noting that N3XT MCP doesn’t replace our existing API infrastructure, but depends on it. We needed a mature and robust API infrastructure to already be in place before we could create N3XT MCP. Our APIs (which you can read about here) remain core to managing access to live account data and workflows.

Where the Model Context Protocol (MCP) changes things is in how AI models interact with this.

First, APIs need developer-written, hardcoded integrations for every use case. MCP, on the other hand, acts as a standard interface that allows AI models to discover and use our banking tools and data in real time. In this way, AI assistants remove friction because they have the ability to query the system, and start approved workflows more easily.

Second, Robotic Process Automation (RPA) automates repetitive tasks through defined rules, but breaks the moment it encounters anything unexpected. Banking interactions with MCP, meanwhile, allows real-time context so the model can reason through complex tasks without relying on fixed logic. A user might ask a multi-step question, or ask the AI assistant to consider bank data in context with another data set that would typically sit outside of view from the bank.

Instead of building custom integrations for every new AI assistant or tool in the enterprise stack, MCP provides a unified standard. You expose the capability once through MCP, and any compliant AI model can securely interact with it.

Ultimately, our APIs provide the execution engine, but MCP provides the language to let AI assistants safely and accurately operate with the environment in real time.

The platform provides governed read-and-write capabilities, allowing AI assistants to analyse transactions, reconcile activity and prepare payments. What can an AI agent do independently today, and which actions must still pass through human approval?

For treasury management and banking, speed matters, but safety and compliance are non-negotiable. N3XT MCP acts as a governed guardrail, ensuring that the same safeguards designed for humans prevent an AI agent from taking unapproved actions.

Here is how that balance works in practice:

AI agents inherit the access permissions of their user. If you only have visibility into a small group of wallets, the agents you build would have that same visibility. So the agent can operate and conduct analysis only within their area of access.

More specifically, an agent can monitor and analyze live account feeds to evaluate cash positions and understand context across different data sources, and automatically match incoming payments against invoices, flag errors, and identify discrepancies. These are all things an AI agent can do independently today.

However, when action moves from preparing data to making payments, an additional layer of governance and permissions ensure that agentic actions comply with existing maker / checker workflows. In this way, payments that have been set to require secondary approval can be initiated by an agent, but they must go to the secondary human approver for final authorization before any money moves. It’s also worth stating that no agent can change approval policies or governance pathways. That is outside of their scope.

So we allow for maximum autonomy and analysis without the risk of ever compromising existing governance and compliance workflows.

In short, AI can get you 100% of the insights you are looking for, and most of the way towards your payment needs, but when it comes to actually making payments, transfers and moving funds, maker / checker workflows are still there to ensure that every transaction is checked and approved by a human before an action is taken.

Allowing an AI system to interact with a corporate bank account introduces significant security and operational risks. How does N3XT ensure that an agent cannot exceed a user’s permissions, access an unauthorized wallet or initiate an improper transaction?

Security in AI isn’t about trusting the model to behave, it’s about architecting systems so that even if models make mistakes, the system architecture prevents an unauthorized action from executing.

We built N3XT MCP with a Zero Trust philosophy for this reason, so an AI agent can never have a “super-user” key or independent access rights. When a person connects to N3XT MCP, the AI agent inherits that user’s access permissions. If a user does not have permission to view a specific wallet, or draft payments above a certain dollar threshold, the agent has the same restrictions. Period.

N3XT MCP also offers a limited set of capabilities to AI agents. Changing permissions is not one of those capabilities. In fact, it’s not even possible via API to change permissions and workflows. This makes it impossible for an agent to ever make a change in this area.

Lastly, rules like maker / checker workflows are enforced at the wallet level, not the user level. This means an agent can never bypass required secondary human sign-offs.

N3XT says its existing maker-and-approver workflows remain in place when clients use AI assistants. How do you preserve accountability and a clear audit trail when a financial action may involve an employee, an AI model and several automated systems?

When multiple entities, a human, an AI model, and backend systems, touch a financial transaction, standard API logging isn’t enough. For auditability, we need to know not just what happened, but who initiated it, what the AI reasoned, and who authorized it.

We maintain absolute accountability by ensuring every request from N3XT MCP includes a tag that links the human user’s session, the specific AI interaction ID, and the backend tool call. If an AI agent drafts a payment, we log which employee issued the prompt, the session, and the tools used by the AI model. There are no anonymous actions in our logs.

When an AI agent acts as the “maker” by preparing a payment, it cannot self-authorize execution. The prepared transaction is pushed into the bank’s standard maker / checker queue. When the human “checker” reviews and approves the payout, their personal authentication token signs the final action. Accountability is preserved.

Which initial use cases are generating the strongest interest from corporate treasury teams and trading organizations: reconciliation, liquidity monitoring, discrepancy detection, payment preparation, reporting or something else?

Every finance team wants end-to-end automation, but despite that, corporate treasurers and trading desks are pretty pragmatic. No one wants to start with the complex and risky workflows; they start where their operational pain is highest and risk is lowest.

Right now, the biggest demand is for reporting. Treasury teams are already swamped in data, and it’s fragmented across multiple banks and partners, so it’s hard to rationalize. They are already using AI assistants to help them see the big picture, but right now, they need to log in to different portals to download positions and statements. Now with N3XT MCP, there is no login needed for N3XT and they can instead have a conversation with their positions.

We expect Payment Prep to be the next use case. We’re already seeing initial excitement for this, and I expect we will see a lot of payment flow creation in the near-term.

N3XT operates as a full-reserve narrow bank that does not lend and backs deposits one-to-one with cash or short-term U.S. Treasuries. Why is this model particularly suited to programmable payments and AI-driven financial operations, and how should clients evaluate its protections compared with conventional FDIC-insured banking?

AI is an accelerator for finance, but the answer to this question isn’t just about AI, it’s about settlement. Traditional banks were built for a world that moved slowly. They rely on a multi-day float to manage and profit from the differences that exist between operating deposits and the bank’s commercial loans.

Introducing real-time, 24/7 settlement requirements, whether it’s introduced by a person or an AI agent, exposes a flaw in the fractional reserve system: to settle funds instantly, you need to have the funds available. In a world where payment velocity is increasing, banks need an equal or greater increase in reserves to make sure those funds are available.

We believe that 24/7 settlement cannot safely coexist with long-term debt issuance on a single balance sheet. Our narrow bank model separates the two and ensures we remain liquid, fully backed, and isolated from the credit risks of a lending desk.

When corporate treasurers compare our narrow bank full-reserve model to FDIC insurance, they should consider how the “safety” is delivered. FDIC insurance caps out at $250,000. For enterprises moving millions, that leaves almost all of their operational capital exposed to the bank’s underlying lending and balance sheet risks.

Full-Reserve Narrow Banking doesn’t rely on insurance at all because we don’t lend. Whether your balance is $100,000 or $100 million, we never lend out your capital so you know it will be there to support your operations and payment needs. We believe this is what’s necessary for a 24/7 instant settlement economy.

N3XT has also introduced the N3XT Digital Dollar, a bank-issued tokenized deposit designed for round-the-clock settlement. How will N3XT MCP interact with tokenized deposits, stablecoins and traditional U.S. dollar payment rails within the same treasury workflow?

First, it’s worth clarifying a point about our model. N3XT is purpose-built to support 24/7 B2B payments with atomic settlement. Legacy rails weren’t designed for atomic settlement, so don’t interact there. That was a deliberate choice.

We spent two years building a modern, blockchain-based core banking system. This includes a private permissioned chain where clients transact in dollars to make payments with other N3XT clients on the network, and public chain access where many of our clients are already transacting. The public chain is where they can transact using the N3XT Digital Dollar (NDD).

MCP allows AI assistants to orchestrate workflows between these two environments. For example, checking NDD balances in a client’s public wallets, and then running a sweep between the wallets if needed, or bridging funds from a private wallet to a public NDD wallet — all while adhering to governance workflows.

So, MCP provides a governed interface to operate across N3XT’s 24/7 digital architecture, both for USD and NDD.

Much of the value of an open standard depends on interoperability. Which AI assistants, enterprise systems and treasury platforms can currently connect to N3XT MCP, and how are you preventing clients from becoming dependent on one model provider or proprietary interface?

The reason we built on the Model Context Protocol (MCP) instead of releasing an SDK of our own was to enable interoperability. Our clients use the tools they use, and in the age of AI, they might even switch more often than they did before.

Because MCP is an open specification, N3XT MCP plugs directly into whichever AI host environment a client already trusts, like Cursor, Anthropic, OpenAI, or Gemini. It also offers native compatibility with orchestration frameworks like LangChain and AutoGen. When it comes to Enterprise systems like ERPs, if those systems have built native MCP connections, then clients can also work across platforms from within their chosen AI platform.

So with MCP, we are giving clients more freedom. If they decide to switch AI providers, or they want to change models to whatever is the latest and greatest, they don’t need to rebuild any connectors. They simply point the new AI model at the N3XT MCP server and they are off and running.

N3XT describes this launch as an early step toward autonomous corporate capital management. How autonomous should corporate finance realistically become, and what technical, regulatory and cultural barriers must be resolved before businesses allow AI agents to manage meaningful amounts of capital?

The goal of autonomous finance isn’t to create a “set it and forget it” black box that moves money without human oversight. Unmonitored autonomy isn’t innovation; it’s a liability.

More realistically, corporate finance should evolve toward engaged autonomy: AI agents doing the data analysis, monitoring, and workflows, while business leaders and finance teams shift from manual execution to policy setting, strategy and approvals.

To grow enterprise trust and give AI agents access and governed control over payments and operating capital, the issues of identity and accountability need to be resolved.

Who is liable if an AI model misinterprets an invoice and triggers an improper payout? Our maker / checker model works to prevent this from happening.

Culturally, we are still early in the transition to agentic finance. As AI takes on larger operational roles, I believe agentic identity will become a topic of increasing importance and focus because trust — in people or in AI — requires accountability.

Thank you for the great interview, readers who wish to learn more should visit N3XT.