For about a decade, design was the biggest game in banking.
It has been all about who has the best app, the best onboarding flow, the smoothest mobile payments experience and the better notifications. The banks that invested in digital experience pulled away from the ones that didn't. Customers voted with their phones, Neobanks grew, and the incumbents who moved built real advantages. But this is no longer enough.
I've been in enough boardrooms over the last twenty years to know this: Banks that have genuinely good digital products - award-winning apps, high NPS, strong digital adoption - are still struggling to scale. Costs keep climbing while operations stay manual. The app is beautiful, but the business operation under the hood isn't, and this is the primary reason AI stays stuck in pilots.
I think what we are living through now is the innovator's dilemma arriving for banking. Banks need to be willing to question everything they built from the ground up: processes, organizational structure, and technology decisions taken over the last few decades. Mobile was a new channel - you replaced the mouse with your finger. AI is different. It doesn't add a channel. It interrogates the operating model underneath every channel you already have.
I want to dive into this and share some of what I see out there.
Here’s the opening slide from my keynote at our recent flagship event where I stood on a stage (with a broken leg) in front of 200+ banking leaders to explain why we are launching the AI-native Banking OS and pivoting to go all in on agentic banking.
Think of banking transformation in 3 levels:
Level 1 is the digital experience.
This level assesses how your app looks and how customers interact with your brand, covering onboarding, mobile banking and self-service journeys. Most banks have done serious work here over the past decade. We have to continue doing this important work, but the market demands more now.
Level 2 is the Unified Frontline.
This is how the frontline business runs. It is the orchestration between customers, employees, and now AI agents across digital channels, the front office, and operations. This layer shifts focus from how the app looks to how the work flows.
Level 3 is Elastic Operations.
Elastic Operations is how the frontline scales. Banks grow volume without growing headcount linearly. AI agents join the workforce without adding operational risk. The result: double the bank without doubling the cost base.
Most banks are solidly at Level 1, and some are beginning Level 2. Very few have reached Level 3.
Only one in four banks worldwide is using AI to gain a real competitive advantage. It’s because they face a serious structural problem in the form of an architecture incapable of scaling AI. Because of this, we are now entering an era where banks’ real competition has shifted to their operations and the architecture it runs on.
To be clear, we are not talking about AI models here - everyone has access to the same models (beside the ones who have access to Mythos).
The whitespace problem banks have
Most banks lack the infrastructure to scale AI. Legacy systems, fragmented data and insufficient orchestration capabilities are blocking real-time, integrated AI applications and agentic workflows. This needs to be solved before we should even think about adding AI.
We need to ask ourselves: When a customer interaction starts in the mobile app and can't be resolved automatically, what happens next? This spans, for instance, a disputed transaction, a loan exception, a complex servicing request, or just a general enquiry about their account.
In most banks, this falls into the whitespace between systems. In response, an operations employee opens four or five systems. They gather context manually, check policy in a separate tool and copy information between screens.
They end up making a phone call to a colleague or they send an email to escalate the matter. I once had a customer tell me their operations desk requested a third physical monitor because two screens weren't enough to fit all the open applications.
That’s the reality of what’s going on in most big banks today, and it's the real cost structure, not the systems. Roughly 60% of banking operations work lives in the whitespace between the systems and not inside them. It lives in the handoffs, exceptions and manual coordination that humans perform because no system owns.
Retail banks could unlock more than $370 billion annually in additional profits by 2030 through large-scale deployment of AI - but only if the underlying operating model can absorb it. For most, that's the part that isn't ready yet.
The reason this matters so much right now is that AI makes the whitespace problem worse before it can make it better. If you deploy agents into a fragmented architecture, those agents are reasoning on incomplete data, operating without shared context, and writing decisions back to different systems. You don't get automation at that point - you just get fragmentation running at higher speed.
Level 2 requires a fundamentally different approach than building great digital experiences. AI is the accelerant, but the operating model is the prerequisite to extract the value.
Moving to the Unified Frontline requires an operating model shift.
What does Level 2 look like in practice?
A customer raises a transaction dispute in the mobile app. Instead of falling into a queue, it triggers a coordinated front-to-back process, with the context already assembled. Transaction history, behavioral signals, prior interactions, policy version in effect at the time are all there, ready and loaded.
An AI agent categorizes the dispute, flags confidence level and prepares the recommendation for the employee who needs to handle the case. A human then proceeds to review and approve. Provisional credit is issued, the evidence gets collected in the background and a back-office analyst resolves the case. This all happens while an authority layer runs a compliance sweep before the case is closed.
This entire journey, start to finish, takes minutes instead of days. All is done by the same actors and the same underlying systems. It’s just a completely different operating model.
BUT it’s really not about AI here. The operating system sitting above those systems is what’s important - coordinating the work, maintaining shared context, governing every action along the journey and connecting everything front to back.
The banks making the most progress here tend to have a strong alliance between the chief digital officer and the chief operating officer. The CDO owns the channel where intent starts. The COO owns the operations where resolution happens. When those two are working toward the same outcome, the transformation becomes surprisingly tractable.
The three things that have to be in place for this to work.
If your team is evaluating what a transformation like this requires, here’s where to start: Shared context and governed authority are the two prerequisites, and front-to-back orchestration connects them to the cost structure.
Shared context comes first
Every system in your bank holds a fragment of the customer truth. The core knows the balance, the CRM knows the relationship, the LMS knows the loan and the fraud system knows the risk signals. None of them knows what the others know - and none knows what's happening right now across this specific customer's current interaction somewhere else in the bank.
You need a semantic layer that assembles that context in real time and makes it available to every actor - human, workflow, and agent - all working from the same truth.
What we've built with Nexus, the Semantic Layer in our Banking OS, does this through a banking ontology - a customer state graph that aggregates data from across the bank into one coherent operational picture. It also stores the decisions made throughout a customer journey - not just the outcomes, but the decision itself, the data that drove it, the policy that governed it. That context travels with the customer across every interaction, creating a complete, auditable record that regulators will ask for as agentic AI scales.
Governed authority comes second
When agents start acting on behalf of customers and employees, the first question every regulator and every risk team asks is: Who authorized that? Under what policy? With what constraints? The answer cannot be "we think the model got it right."
Every action needs a verifiable chain of authority and a record that captures what policy applied, what actor executed it, what model version was used, and what the outcome was. Without that, AI can’t scale in a heavily regulated environment like banking. It gets constrained or shut down entirely. This is a good thing, because this is why people trust banks - so it needs to be done right.
The way we've approached this with Sentinel, the Authority Layer in our Banking OS, is through a simple principle: just as you need to know your customer and know your employee, you need to know your agent. Where is it registered? What is it authorized to do? What policies govern its behavior in a particular subdomain? Every action it takes produces a Decision Token - a traceable record of the policy applied, the actor identity, the model version, and the full context of the decision. Without this structure, I don't see a future for AI at scale in banking. Governance needs to be designed into the execution layer from day one, not added as a check at the end.
Front-to-back orchestration is what connects both of those to actual cost reduction.
Bank-work lives in the coordination between channels, the front office and operations. An operating model that only covers the digital layer leaves 60% of the cost structure untouched and unaccounted for. The orchestration has to run the full length of the process - from the customer interaction that starts it to the back-office resolution that closes it. Unfortunately, you can’t go buy this off a shelf. It’s an operating model to build progressively, one domain at a time.
Where to start unifying the frontline?
Start with an MRI scan of your customer journeys. Map intent to resolution across your key workflows and measure the resolution time. Anything that takes longer than a few minutes to resolve is almost certainly carrying significant manual coordination cost in the background. Those are your highest-friction, highest-value starting points. Pick one - transaction disputes, loan origination exceptions, KYC remediation - whatever has the highest manual volume and the worst resolution time in your operation.
Map the current process and count the systems involved. Do not forget to count the handoffs and measure how long it takes end to end.
Next, build the coordinated version, front-to-back, with shared context and governed authority, and run it alongside the existing process. Then, go and measure the difference. Come back and tell me what you find and we have a discussion about it.
We did this with dispute resolution across several implementations.
Before, the banks had six to seven different systems, more than twelve manual steps, three teams involved, and an average resolution of around 20 minutes.
After: one workspace, one end-to-end orchestrated process, and a process as quick as two minutes. That’s one (of many) concrete use cases you can deploy.
Once one domain works this way, the pattern is established and you can move onto the next. The architecture is in place, the agents are trained, and the best part is the next domain deploys faster, and the one after that is even faster.
AI agents currently represent 17% of AI-derived value across industries - BCG expects that to reach 29% by 2028. The infrastructure built now is what makes that value accessible when it arrives in the very near future.
The practical question to take back.
For ten years, the question driving technology investment was: how good is our digital experience?
That question still matters, but it's no longer the only one worth asking.
The questions I'd add alongside it are:
How well does our frontline run right now?
How fast do we resolve exceptions?
How many systems does one employee touch to close a single case?
Does our cost-to-serve stay flat as volume grows - or does it scale with headcount?
Those are Level 2 questions. They don't require a transformation program to start answering. They require picking one process, mapping it out and building the coordinated version with agentic workflows in the mix.
Frequently asked questions
Why do AI agents fail in banking?
AI agents fail in banking because the foundation underneath them is fragmented, not because the models are wrong. When agents are deployed across disconnected systems, they reason on incomplete data, operate without shared customer context, and write decisions back to different systems. The result isn't automation - it's fragmentation running at higher speed. Around 60% of frontline banking work lives in the whitespace between systems, and that's where agents fail.
What is the whitespace problem in banking?
The whitespace problem is the operational work that happens between banking systems - the handoffs, exceptions, manual coordination, and copy-pasting that no individual system owns. When a customer interaction can't be resolved straight-through, it falls into this whitespace: employees open four or five systems, gather context manually, check policy in a separate tool, and coordinate across teams. That coordination overhead is where most operational cost, delay, and risk accumulates.
What is agentic banking?
Agentic banking is the operating model where AI agents join customers and employees as active participants in banking work - not as tools that assist, but as actors that execute. They categorize cases, prepare recommendations, gather evidence, and resolve routine tasks autonomously within defined guardrails. The shift from AI as a feature to AI as a participant in the frontline is what distinguishes agentic banking from earlier automation approaches.
What does a bank need before deploying AI agents at scale?
Two things, in order. First, shared context: a semantic layer that gives every agent, employee, and workflow the same real-time view of the customer - the same balance, the same relationship history, the same active cases, the same policy version. Without it, agents reason on fragments. Second, governed authority: a mechanism that records who authorized each agent action, under what policy, with what model version, and with what outcome. Without that audit trail, AI can't survive regulatory scrutiny in banking.
How should a bank start building a Unified Frontline?
Start with an MRI scan of your customer journeys. Map intent to resolution across your key workflows and measure resolution time. Anything taking longer than a few minutes is carrying significant manual coordination cost. Pick one high-friction journey - transaction disputes, loan origination exceptions, KYC remediation - and build the coordinated version front-to-back with shared context and governed authority. Measure the difference against the existing process, then move to the next domain. Each successive domain deploys faster because the architecture is already in place.
