Aurélien Bonnel is a seasoned financial technology executive and engineer who has spent over 14 years architecting the backbone of modern finance. From developing high-frequency trading systems at Deutsche Bank to building the blockchain-based Signet platform for Signature Bank, his career has been a masterclass in secure, real-time infrastructure. Now, as the CTO and Founder of N3XT, he is leading a shift toward a new era of banking where 24/7 settlement and artificial intelligence converge. By combining a narrow banking model with the innovative Model Context Protocol, he is redefining how corporate treasuries manage capital in an increasingly automated world.
In this conversation, we explore the systemic inefficiencies of traditional banking that inspired the creation of a full-reserve institution and the technical nuances of connecting AI agents to live financial workflows. We discuss the critical role of governed “maker/checker” protocols in maintaining security, the transition from fragmented reporting to conversational data analysis, and why the future of finance relies on “engaged autonomy” rather than black-box automation.
Traditional banking often involves settlement delays that benefit the institution rather than the client. Having worked in high-frequency trading and blockchain, what were the specific friction points that convinced you the industry needed a narrow bank built from the ground up?
The reality of traditional banking is often hidden in plain sight, and you see it most clearly when a payment sent on a Friday doesn’t actually land until Tuesday. During my time at Deutsche Bank and later while building blockchain infrastructure, I kept seeing this persistent gap between “Available Balance” and the actual settled funds that a business can use. It became clear to me that these delays aren’t a technological limitation, as the tools for instant settlement have existed for years, but rather a deliberate part of the legacy business model. Banks profit from the interest earned on that “float” while money sits in limbo, which essentially means they are making money off their clients’ lack of liquidity. This realization convinced me that you cannot simply patch old systems; you have to build a narrow foundation where the bank does not lend out customer deposits. By backing every dollar one-to-one with cash or short-term U.S. Treasuries, we can finally align the bank’s architecture with the need for instant, 24/7 settlement.
The introduction of N3XT MCP marks a shift in how AI interacts with financial data. Could you explain what the Model Context Protocol offers that makes it superior to traditional APIs or standard robotic process automation for corporate finance?
It is important to understand that N3XT MCP doesn’t replace our existing API infrastructure; rather, it sits on top of it as a sophisticated translation layer. While traditional APIs are excellent for execution, they require developers to hardcode integrations for every single specific use case, which creates a lot of rigid friction. Robotic Process Automation, or RPA, is even more fragile because it relies on fixed rules and tends to break the second it encounters an unexpected variable or a slight change in a UI. The Model Context Protocol changes the game by providing a standard interface that allows AI models to actually discover and reason through banking tools and data in real time. Instead of following a script, the AI can understand the context of a multi-step query, such as analyzing bank data alongside an external dataset that the bank wouldn’t normally see. This flexibility means we can expose a capability once through MCP, and any compliant AI model can then interact with it securely without us needing to build a custom connector for every new tool.
When we talk about AI agents analyzing transactions and preparing payments, there is a fine line between autonomy and risk. How do you define the boundary between what an agent can do on its own and what still requires a human hand?
In the world of corporate treasury, speed is a huge advantage, but safety is absolutely non-negotiable, so we’ve built N3XT MCP with very specific governed guardrails. Today, an AI agent can independently handle the “heavy lifting” of data analysis, such as monitoring live account feeds, evaluating cash positions across multiple sources, and even matching incoming payments against invoices to flag discrepancies. These are analytical tasks where AI excels because it can process vast amounts of information without getting tired or missing a line item. However, the moment a task moves from preparation to the actual movement of funds, our “maker/checker” workflow kicks in as a hard stop. An AI agent can act as the “maker” by drafting a payment based on its analysis, but it can never be the “checker” that authorizes the final execution. This ensures that every single dollar moved is ultimately approved by a human who has the final say, preserving the same compliance standards that have protected financial institutions for decades.
Security is the primary concern for any treasurer considering AI integration. How does the N3XT architecture prevent an AI agent from overstepping its bounds or accessing sensitive wallets it wasn’t meant to see?
Our approach to security is rooted in a Zero Trust philosophy, which means we never assume an AI model will behave perfectly; instead, we build the system so that the model physically cannot perform an unauthorized action. We achieve this by ensuring that every AI agent inherits the exact permissions of the human user who is interacting with it. If a specific employee doesn’t have the right to view a particular wallet or initiate payments over a certain dollar threshold, their AI assistant is restricted by those same identical walls. Furthermore, we’ve designed the MCP server so that it doesn’t even have the capability to change permissions or governance pathways via API, making it impossible for an agent to “escalate” its own privileges. By enforcing these rules at the wallet level rather than the user level, we ensure that the required human sign-offs are a structural requirement of the bank’s core, not just a suggestion the AI can bypass.
In a complex workflow involving humans, AI models, and automated backend systems, the audit trail can quickly become murky. What steps have you taken to ensure that every financial action remains fully accountable and transparent?
Accountability is the bedrock of trust in finance, so we realized early on that standard API logging simply wouldn’t be enough for an AI-driven environment. When a transaction is initiated, we need to know the full context: who prompted the AI, what logic the AI used to reason through the task, and who eventually authorized the result. To solve this, every request flowing through N3XT MCP is tagged with a unique ID that links the human user’s session directly to the specific AI interaction and the resulting backend tool call. There are no “anonymous” actions in our system because if an agent drafts a payment, our logs show exactly which employee was behind the keyboard and what tools the model accessed to get there. Because the final execution requires a human checker to sign off with their own personal authentication token, the audit trail remains a clear, unbroken chain of human responsibility.
As you engage with corporate treasury teams and trading desks, which specific pain points are driving the most immediate interest in AI-driven banking tools?
Right now, the most overwhelming demand we see is for reporting and data synthesis because finance teams are currently drowning in fragmented data spread across different portals and partners. Treasurers are tired of logging into five different systems just to download statements and manually reconcile their positions, so the ability to have a “conversation” with their live data is a massive win for them. We are seeing clients use AI assistants to get a high-level view of their liquidity in seconds, rather than spending hours in spreadsheets. Beyond reporting, we are seeing a lot of early excitement for payment preparation, where the AI handles the tedious work of drafting transfers and identifying which invoices are ready to be paid. I expect that in the near term, this “payment flow creation” will become a standard part of the treasury toolkit, as it allows teams to move from manual data entry to a more strategic oversight role.
N3XT operates as a narrow bank, which is a departure from the traditional fractional reserve model. Why is this specific structural choice so important for the success of programmable payments and AI?
The traditional fractional reserve system was built for a world where money moved at the speed of paper, relying on a multi-day “float” to manage the gap between deposits and long-term commercial loans. However, when you introduce AI agents and real-time settlement that operates 24/7, that legacy model starts to show its cracks because you need the actual funds to be available at every moment. We believe that trying to support instant, high-velocity settlement on the same balance sheet as long-term lending is inherently risky, which is why our narrow bank model is so critical. By separating lending from payment operations, we stay fully liquid and isolated from credit risks, ensuring that the capital is always there when a programmable instruction hits the ledger. For an enterprise moving millions of dollars, this is a much more robust form of safety than FDIC insurance, which only covers up to $250,000 and leaves the rest of their operational capital exposed to the bank’s lending risks.
The N3XT Digital Dollar (NDD) allows for settlement on public and private blockchains. How does N3XT MCP help a treasurer manage these different rails and tokenized assets within a single workflow?
We spent two years building a core banking system that could bridge the gap between traditional dollar accounts and the world of tokenized deposits on a blockchain. N3XT MCP acts as the orchestrator for this entire environment, allowing an AI assistant to see across a client’s private wallets on our permissioned chain and their public wallets where they might hold NDD. A treasurer can ask the AI to monitor their public NDD balances and then automatically trigger a “sweep” or a bridge from a private wallet if funds get too low, all within a governed framework. This means the user doesn’t have to worry about the underlying technical complexity of moving between public and private rails; they just describe the desired outcome to the assistant. It essentially turns a complex multi-chain treasury operation into a seamless, conversational experience that still respects every one of the bank’s security protocols.
Interoperability is often the “achilles heel” of new financial technology. How does building on an open standard like MCP protect your clients from being locked into a single AI provider or a proprietary interface?
The decision to build on the Model Context Protocol instead of a proprietary SDK was a very deliberate move to give our clients maximum freedom and flexibility. We know that the AI landscape is moving incredibly fast, and a model that is the market leader today might be overtaken by a competitor in six months. Because MCP is an open specification, our system plugs directly into the host environments that clients already use and trust, whether that is Anthropic, OpenAI, Gemini, or even developer tools like Cursor. If a company decides to switch their primary AI provider, they don’t have to rebuild any of their banking connectors or integrations; they simply point the new model at the N3XT MCP server. This “plug-and-play” compatibility also extends to orchestration frameworks like LangChain, ensuring that N3XT remains a flexible part of the broader enterprise tech stack rather than a siloed application.
As we move toward a future of more autonomous capital management, what do you see as the final hurdles—whether they are cultural, regulatory, or technical—that businesses need to clear?
I think the biggest hurdle is a cultural one: moving away from the idea that autonomy means a “black box” that operates without oversight. Real innovation in corporate finance isn’t about setting an AI loose to move money unmonitored; it’s about what I call “engaged autonomy,” where the AI handles the heavy operational lift while humans focus on strategy and policy. We also have to solve the looming questions of agentic identity and accountability, such as who is liable if an AI misinterprets a complex invoice and suggests an improper payout. This is exactly why we lean so heavily on our maker/checker model, as it provides a clear framework for trust by keeping a human in the loop for the most critical actions. As these systems become more common, I believe we will see a shift where trust is built not just through personal relationships, but through the transparent, auditable architecture of the systems we use to manage our wealth.
What is your forecast for the future of AI in the banking sector over the next five years?
In the next five years, I expect the very definition of a “bank account” to change from a passive storage bin for capital into an active, intelligent partner that manages its own liquidity. We will see the total disappearance of manual reconciliation for mid-to-large enterprises, as AI agents will handle the real-time matching of every transaction as it happens, 24/7. However, this transition will also force a massive reckoning for traditional fractional-reserve banks, as the demand for instant settlement from these AI agents will make the old “multi-day float” model completely unsustainable. Eventually, I believe the industry will split into two distinct tiers: high-velocity narrow banks that provide the infrastructure for automated commerce, and traditional institutions that focus almost exclusively on long-term credit and lending. The winners will be the ones who realize that in an AI-driven economy, the most valuable asset a bank can offer isn’t just money—it’s the programmable, instant certainty of that money.