The bank with no interface

The best bank interface is the one you never see. That sounds like a provocation. It's not. It's a design conclusion - and it follows directly from watching how people actually interact with their company's finances today.

In the previous memo, we argued that the next great fintech won't be a bank - it'll be an application layer on top of blockchain rails, built for AI agents. This piece is about what that means for the thing everyone assumes a bank needs: an interface. The answer is surprising.

Bank of the future is a ledger, a memory layer and a set of agents. The interface lives somewhere else entirely.

Who actually uses a corporate bank?

Start with the people. A company's financial stack is touched by a surprisingly diverse cast of characters, each with different needs, different contexts and - critically - different tools they already live in.

• Founders and owners. In smaller companies, they have direct access. In larger ones, shareholders are abstracted away and top management operates on their behalf. Either way, they care about the big picture - cash position, runway, risk.
• The CFO. Full access, full accountability. A regulated, auditable role. The person who needs to understand cash flow, compliance with plan and where every dollar sits across every account. In a large company, they're supported by a small army: financial controllers who verify that every purchase matches every contract, FP&A teams who build the models, treasury teams who ensure the right amount of money is in the right account in the right currency at the right moment.
• Employees. They make purchases on behalf of the business, go on work trips, file expense reports, chase down receipts. Banking is the least interesting part of their job - and they treat it accordingly.
• Managers. They sit in between. They have budgets, they approve expenses, they check that spending aligns with what was planned. Systems. Shopify stores, Stripe integrations, payroll providers, ERP platforms - automated connections that move money in and out without a human touching anything.

That's a lot of people. Now ask a different question: where do these people actually spend their time?

Nobody lives in the bank

Here's the thing that should reshape how we think about banking UX: not a single one of these people wants to be in a banking app. And the data proves it — the average NPS for U.S. banking websites is negative 17%. More detractors than promoters. That's not a product that needs improvement. That's a product category that needs to be rethought.

Employees open the banking app once every three thousand years. They forget every password, every login, every security question. And when they finally get in, it's to do something painful - dig up receipts scattered across their phone camera roll, their email, some random messaging thread, and somehow upload all of it into an interface they use so rarely that every button feels foreign. 60% of business travellers say expense submission is the single worst part of business travel. 53% say it's worse than doing their taxes. These same people spend hours a day in ChatGPT and Claude for actual work. The banking app is the absolute last place they want to be.

And the receipts - Christ, the receipts. They're in your email. They're in your photos. They're in some PDF a vendor sent on WhatsApp. You're supposed to collect all of this, organise it, submit it through an app you barely remember exists. 49% of companies list lost receipts as a top-three expense management pain point. Each expense report costs $58 to process and takes 20 minutes of company time - and 19% of them contain errors that cost another $52 each to fix. The most natural thing in the world would be to just snap a photo and send it from wherever you already are - iMessage, Slack, whatever. Not open yet another app that you need for ten minutes a quarter.

Financial controllers live in Excel and their accounting system. That's their natural habitat. Always has been. 96% of FP&A professionals use spreadsheets for planning at least weekly. 89% of finance teams rely on Excel even when they already have dedicated planning software. 82% of finance professionals report an emotional attachment to the tool. They might log into the bank to push a payment through, but the actual work - the verification, the reconciliation, the cross-referencing - happens in spreadsheets. It's always happened in spreadsheets. No banking dashboard has ever come close to replacing Excel, and no banking dashboard ever will. The bank is a waypoint, not a workspace.

CFOs and treasury teams - same story. They need to see consolidated positions across a dozen banks, multiple currencies, various time horizons. Large U.S. companies average 26.4 banking relationships. One in four companies manages six to ten banks. Each relationship means a separate portal, separate credentials, separate data formats. 58% of treasurers rank visibility into global cash positions as their number one challenge - and it's a challenge specifically because the data is fragmented across all those portals. The last thing on earth they want is to log into thirty-three separate banking interfaces and click around in yet another clunky UI. They build everything in Excel. They always have.

Every single role follows the same pattern. Their primary workspace is somewhere else. The bank is something they visit reluctantly, briefly and as infrequently as possible. Up to 20% of corporate clients switch their primary bank in any given year — and when asked why, they cite wanting a user-friendly platform and easier API integrations above product features. They're not asking for a better interface. They're asking to not need one.

Banks have traditionally measured engagement by transaction volume, not by time spent in the app. That turns out to be prophetic. The interface was never the product. The ledger was.

Match the interface to the task

There's a basic principle in interaction design that banking has been violating for decades: the interface should match the nature of the work.

Tabular data demands spreadsheets. When you're working with transaction lists, cash flow projections, budget comparisons - you need to sort, filter, pivot, add columns, build formulas. You need to play with the data. Excel does this. Nothing else comes close. Finance teams using spreadsheets spend 80% of their time collecting data and just 20% finding insights - and even that ratio beats what banking portals offer, which is closer to 100% collection, 0% insight. The best analytics tool for any business has always been a spreadsheet. The best way to improve it isn't to rebuild it - it's to give it a brain. An AI co-pilot inside Excel that can pull live data through an MCP connection to your bank, build the charts that used to take thirty clicks and twelve formulas, model scenarios on the fly, and pull additional context through APIs. You stay in your natural habitat. The bank feeds it.

Commands and orchestration demand conversation. Here's what the ideal workflow actually looks like: you're in Claude or ChatGPT, connected to everything - your bank, your accounting system, your email, your contract management tool. You say: "Find all invoices from last month. Check what we actually agreed to in the contracts - they're in PandaDoc. Compare. Flag anything that doesn't match. For everything that checks out, schedule payment for Thursday." One command center. The bank is one rail among many that the agent can access. You don't go to the bank to do this. You go to your command center - the AI - and the bank is just one system it orchestrates. 59% of finance functions already use AI, up from 37% in 2023. 58% of finance employees use AI tools at least a few times per year - the second-highest adoption rate of any industry. This isn't a future prediction. It's already where these people spend their time.

Approvals demand structured workflow. Sign this payment. Review this expense. Confirm this transfer. These are high-stakes, discrete actions that need clear presentation, audit trails, and deliberate confirmation. This is where the bank's own interface still has a role - but a much narrower one than today.

Corporate banking involves at least three fundamentally different types of work, and no single interface serves all of them well. The current approach - cramming everything into a banking portal - serves none of them well.

The bank becomes invisible infrastructure

So strip it down. What does a bank actually need to provide?

• A universal ledger. The single source of truth for all financial states - balances, transactions, positions across every account, every currency, every entity. This is the core. It doesn't need a pretty interface. It needs to be accurate, real-time, and accessible through APIs and agent protocols.
• A memory layer. Context that agents need to operate - rules, policies, approval workflows, historical patterns, business-specific logic. This is what allows agents to make decisions on behalf of the company.
• Payment rails. The ability to actually move money - increasingly through stablecoin infrastructure that settles in seconds rather than days.
• An agent platform. The infrastructure for financial agents to operate, collaborate and execute workflows. One agent creates a payment. Another checks compliance. A third optimises timing and routing. Each company configures its own processes and rules.

The bank is a ledger and a memory. Not a destination. Not an interface. Infrastructure that integrates with everything and presents itself through nothing.

Where the interaction actually moves

• To spreadsheets - for everything analytical. Your Excel connects directly to the bank's ledger through an MCP integration. Live data flows in. You build your analysis, ask Copilot or Gemini to generate charts, model scenarios, run comparisons - without leaving the sheet. You don't go to the bank to understand your finances. You pull your finances into the tool where you've always understood them.
• To AI chat interfaces - for everything operational. Claude, ChatGPT, Gemini - these become the command center. Connected to your bank, your accounting system, your email, your document storage. The AI orchestrates across all of them. Payments used to be something you could only initiate inside the bank - through a clunky form, in a clunky app. Now you can create them from a chat message. On demand. On the fly. 87% of CFOs say AI will be extremely or very important to their finance operations by 2026. Gartner predicts 90% of finance functions will deploy at least one AI-enabled solution by then. The migration is already underway.
• To existing tools - for everything contextual. An employee photographs a receipt and sends it through the messaging app they already have open. A manager approves an expense inside their project management tool. An ERP reconciles incoming payments automatically. The bank is always present as infrastructure. Never as the interface.

Your bank's biggest competitor isn't another bank. It's a Claude chat with an MCP plugin.

What remains in the bank's own UI

If analysis moves to Excel and commands move to chat, what's left?

A much smaller, much more focused application. The bank's own interface becomes an administration and governance layer - not the place you go to look at transactions and charts.

• Tasks and approvals. The daily activity. Reviewing and signing transactions, working through approval queues. This is the primary reason anyone opens the bank, and probably the only daily use case.
• Team management. Onboarding employees, setting roles and permissions, managing who can see and do what. Inviting a new team member through email or through an HR integration. Standard administration.
• Agent management. This is the new surface. Connecting external agents - Claude Code, ChatGPT, custom systems. Configuring internal agents. Setting rules and policies for automated operations. Defining what agents can do, what requires human approval, what escalation paths look like. This is where the agentic architecture becomes tangible for users.
• Integration management. Connecting accounting systems, ERP platforms, payment providers. Configuring how data flows between everything. Setting up the MCP connections that let external tools pull bank data in real time.
• Compliance and audit. KYB/KYC, regulatory requirements, audit trails. Things that need a controlled, purpose-built environment.

That's it. The bank interface becomes an admin panel for governing a financial system - not the system itself.

People used to go to the bank to view transactions because there was no other convenient way to see them. They used to initiate payments there because there was no other way to create them. Both of those constraints are gone. Transactions flow into whatever spreadsheet or tool you want. Payments can be created from any connected interface. What remains is governance - and governance is what the bank should focus on.

Generative UI: when you do need the bank

There's one more layer worth considering. Some people, some of the time, will want to interact with the bank directly. Maybe they want to explore something specific. Maybe their company runs an in-house system that doesn't integrate with Claude or ChatGPT. Maybe they just prefer it.

For this, the bank needs an interface - but not a traditional one. In an AI-native world, the interface doesn't have to be static. Traditional banks release new features every four to six months. Fintechs release every two to four weeks. The average age of a universal bank's IT applications is fourteen years. You can't win the interface game on that infrastructure. But you can skip it entirely.

Call it Generative UI. An interface that builds itself on the fly from a set of primitives - tables, cards, charts, approval forms - assembled dynamically by an AI layer in response to what you need right now. "Show me all transactions over $10,000 this month" - a filtered, sortable table materialises. "Show me account balances across all entities" - a card layout appears with real-time data. "Compare our FX exposure this quarter versus last" - the visualisation builds itself.

This isn't a dashboard you configure once. It's an interface that materialises when you need it and dissolves when you don't. No one navigates through seventeen levels of menus to find the right report. You ask, and it appears.

What makes it work. The system understands three layers of context simultaneously:

• First, intent - what are you trying to do right now?
• Second, memory - are you a CFO who usually works with cash flow, or a controller in the middle of budget planning? What did you look at in your last ten sessions?
• Third, external context - time of day, which entity you're working with, which of your multiple subsidiaries is relevant. All of this feeds the generation layer. The interface doesn't just respond to your query. It anticipates what you're likely to need.

Why it's better. You don't need to learn the interface. The interface learns you. It explains itself, adapts to your request, and tailors itself maximally to what you're trying to accomplish. From the builder's perspective, you're shipping primitives and composition rules - not hand-crafted screens that need to be manually updated every time something changes. It's infinitely customisable per client. Any table, any join, any slice of data, any visualisation the client wants - generated on demand. And it carries context and memory that previously required manual configuration or didn't exist at all.

What it probably looks like. Think about what finance people are actually comfortable with: spreadsheets. A canvas where you can play with data, manipulate it, build tables on the fly. The natural form factor for Generative UI in banking is probably a large canvas - part chat, part spreadsheet - where you can ask questions and the system generates interactive, editable tables and visualisations directly. You build something, you tweak it, you share it. The views you create become shareable artifacts that other team members can open, modify, and build on. It's collaborative, it's persistent, and it's dynamic - when you come back, your data is fresh, your context is preserved, and your workspace is waiting.

What remains hard. Four real challenges:

• First, speed - consumers expect instant response, and generation can't take several seconds. The primitives need to render immediately, even if the data is still streaming in.
• Second, tabular data - financial interfaces are overwhelmingly tables, and current language models aren't always great with structured tabular operations. This will require specialised models or fine-tuning that handles joins, pivots, and aggregations natively.
• Third, the spreadsheet problem - how do you make generated views feel as fluid and manipulable as a real spreadsheet, with live-updating data, persistence across sessions, and the ability to manually edit and explore?
• Fourth, collaboration - some of these generated views need to be shared, discussed, and worked on together. That means real-time state, permissions, and a shared context layer.

These are solvable problems. Protocols like MCP Apps and AG-UI are already enabling the integration layer. The model capabilities are improving fast. And the underlying technology stack - from streaming UI components to real-time data sync - exists today. The pieces are there. Someone just needs to assemble them for finance.

What this means for how you build a bank

Don't optimise the interface. Optimise the integrations. The quality of the bank is measured by how seamlessly it connects to Excel, to Claude, to accounting systems, to every tool where financial work actually happens. MCP plugins, API quality, real-time data access - this is the product surface that matters. Not the dashboard. Fewer than 50% of banks even offer open APIs to corporate clients today. That's the gap.

Don't compete for attention. Compete for trust. Banks have traditionally wanted users in their app, engaging with their dashboards, using their tools. An AI-native bank does the opposite. It pushes interaction outward, to wherever the user already is. The bank's value isn't in capturing eyeballs - it's in being the most reliable, most accessible financial infrastructure that every other system can depend on.

Don't build features. Build primitives. The Generative UI concept only works if the underlying system is composed of clean, composable building blocks - tables, approval workflows, account views, transaction filters. Ship primitives, not pages. Let the AI layer assemble them.

Design for agents first, humans second. The majority of interactions with the bank will be through agents - AI assistants, automated systems, custom workflows. The human-facing interface matters, but it's secondary. If you get the agent layer right, humans benefit automatically through whatever tool they prefer.

---

The tools people use for work - AI assistants, modern spreadsheets, integrated platforms - have leapt ahead of the tools they use for banking. The gap is widening every month. Finance professionals now spend more time in AI tools than in their banking portals. Corporate banking portals score negative 17% NPS while ChatGPT has 800 million weekly active users. That's not a trend - it's a verdict.

The bank that recognises this - that stops trying to be the interface and starts being the infrastructure - will win. Not by building a better dashboard, but by disappearing into the tools where work already happens.

The best bank interface is the one you never see. If you see the same thing, reach out.