The decade that was, and the decade that isn't
I've been building banking software for over twenty years. In that time, I've watched banks invest billions chasing a digital surface that looked like Revolut's. Better mobile apps. Cleaner onboarding flows. Smoother card experiences. And to be fair - that mattered. Customers noticed. Adoption moved.
But here's what I keep telling banking leaders now: the last ten years, the battle was about user experience and having a beautiful mobile app. The next ten years will be decided by whether banks can coordinate humans and agents across the entire frontline - not just surface them in a single app.
That's a different problem. And banks are still largely solving the wrong one.
Three levels. One reason banks are stuck at level one.
When I talk to banking executives about where they are in this transition, I think about it in three levels.
Level one is the app. You built a mobile banking experience. It works. Customers can check balances, move money, apply for a product. You compete on that surface. Banks reached level one over the past decade, some faster than others. The Revoluts and Nubanks of the world forced the pace, and incumbents followed.
Level two is the operating model. This is where customer centric banking lives - not in the design of the app, but in how the bank executes work behind it. Who handles the exception when a loan application stalls? How does a customer who started a dispute on mobile get a coherent response from the contact center without re-explaining themselves? How does an RM at a regional bank see everything that matters about a client before a meeting, across a dozen disconnected systems? Level two is about the operational reality behind the customer-facing surface. And banks are still deep in manual coordination here.
Level three is full agentic execution. Ten years from now, I believe everything will be agentic - every task, every role executed by agents working alongside people. You need to govern that, control it, make it safe. But level three is not the problem to solve today. The problem today is that banks haven't solved level two.
Fragmentation is what keeps banks trapped at level one. Not ambition. Not budget. The architecture underneath.
Fragmentation kills intelligence
I said this in a recent conversation about where banking is heading: "Fragmentation kills intelligence." That's not a metaphor. It's the operational reality of what happens when you try to deploy AI across a stack built from dozens of disconnected systems.
Banks don't need more channels. Banks don't need more systems - because you increase the fragmentation problem. What we need now is to unify these systems and make sure they can work as one.
Take a specific case: a customer wants to resolve a payment dispute. They open the app, describe the issue, and the AI agent starts handling it. The agent needs to read the transaction record, check the fraud system for associated alerts, verify the customer's history, and draft a resolution. On a fragmented stack, each of those steps hits a different system with a different data model and a different view of who the customer is. The agent either slows to a crawl, fails to act, or - worse - acts on partial information.
McKinsey's research on AI in banking consistently shows that the distance between pilot and production is almost never a model quality problem. It's an architecture problem. The banks that get AI into production fast are the ones that built a unified execution layer underneath.
That's what the AI-native Banking OS is designed to be - not a channel, not a point solution, but the control plane that sits above systems of record and makes everything above the ledger work as one. It connects customers, employees, and AI agents around a single version of what's happening.
What the next decade of customer centric banking looks like
The vision I keep coming back to is one where three actors work together across the bank's entire frontline: customers, employees, and AI agents. Historically banks had two. Growth meant hiring. AI introduces a third actor - and that forces a redesign of how work gets authorized, routed, and completed across the frontline.
Think about what changes when those actors work from the same context. A customer starts an application on mobile. An AI agent prepares the case, checks eligibility, and flags what's missing. An employee picks up exactly where the agent left off, with full context and no duplication. The customer never re-states themselves. The employee never digs through systems. The agent never acts beyond its authorized scope.
That's not a feature. That's an operating model. And BCG's banking CEO surveys show that the leaders pulling ahead in the AI era are the ones who've stopped thinking about AI as a capability to add. They've started thinking about it as a new actor in the frontline operating model.
You can run a bank with 500 people and 5 million customers - but only if the architecture allows agents to absorb the coordination load that humans currently carry. Without that foundation, every agent you add increases the complexity you have to manage manually.
The governance question nobody is asking loudly enough
Here's the part of this conversation that I think gets underweighted. When I talk about agents taking on more and more of the operational load, the immediate reaction from banking leaders is usually about risk: what if the agent does the wrong thing?
That's exactly the right question. And it's why governance isn't a compliance checkbox - it's a core capability of customer centric banking in the AI era.
Every agent action needs authorization. Autonomy should be earned incrementally - measured against outcomes and revocable when something breaks. The human-in-the-loop principle isn't about slowing AI down - it's about knowing, at every step, that a decision was made for the right reason by the right actor under the right authority.
Sentinel, as we've built it inside the Banking OS, does exactly this: no action executes - by any actor, human or agent - without a Decision Token. That token records who acted, under what policy, and what they decided - enough to reconstruct the decision for any regulator. It's what makes AI-driven customer centric banking auditable and not just fast.
Forrester's work on AI governance in financial services makes the case clearly. Regulators aren't asking banks to slow down AI adoption. They're asking banks to prove they know what their AI is doing. Decision Tokens are that proof.
The mid-piece stop: what a unified frontline looks like in practice
Let me walk through what this looks like for a bank making the shift.
Before: A customer calls about a failed international transfer. The contact center agent opens four systems to find the transaction. They check a fifth for compliance flags. They manually copy findings into a case management tool. They ask the customer to confirm details they've already provided twice. Resolution takes three days.
After: The customer messages through the conversational banking interface. An agent reads the Customer State Graph - which already contains the transaction record, the compliance status, and the customer's communication history - and opens a structured case. An employee picks up a pre-populated workspace with all the context they need. Resolution takes four hours. The customer was never asked to repeat themselves.
The difference isn't the AI model. The difference is whether the systems underneath share a common operational model, or whether a human has to bridge those gaps manually. Building agents that work in banking requires that foundation - shared semantics, governed execution, unified state.
That's not a future state. Banks are building this now. And the ones that get there first will set the standard the rest of the industry has to match.
The questions to take back to your own organization
Where does your frontline coordination live? Map one customer journey end to end - not the digital touchpoints, but the operational steps behind them. Count how many systems an employee touches. Count how many times the customer has to re-state context. That number tells you how far you are from a unified frontline.
What percentage of your frontline work lives in the whitespace between systems? Across the 120-plus bank implementations we've worked through, it's close to fifty percent. Exceptions, handoffs, and manual coordination that no system owns. That's where your cost lives. That's also where your biggest AI opportunity lives - if the architecture allows agents to operate there.
Are you solving a level-one problem or a level-two problem? Another app redesign, another channel investment - these are level-one moves. The banks winning the next decade are investing in the operating model: unified execution, shared context, governed agentic capability across the frontline.
When your regulator asks how you govern AI decisions, what do you show them? If the answer is a policy document, that's not enough. You need a technical trail - Decision Tokens, model versions, policy-applied records. AI-native platforms build this into the execution layer, not as an afterthought.
Is your AI strategy a capability strategy or an architecture strategy? Capabilities can be bought. Architecture has to be built - and it takes time. The banks pulling ahead started their architecture work eighteen months ago. The ones still buying point solutions are accumulating the debt they'll have to pay later.
The banks that win the next decade will be the ones that get the operating model right - where customers, employees, and AI agents work from a single source of operational truth. Architecture is destiny. The foundation comes first.
If you want to compare notes on where your organisation sits in this transition, send me a message.
