Fernando Zandona (CEO) and Adrian Congiu (VP of Product)
16 November 2026 • 10 min read
Executive summary
Agentic AI in banking pays off when banks start with one painful process, earn trust with read-only access and build in governance from day one. Mambu’s CEO Fernando Zandona and VP of Product Adrian Congiu explain why waiting is the riskier option, and how to begin.
Introduction
Agentic AI is in active adoption at 52% of the financial services firms surveyed for the 2026 Global AI in Financial Services Report, yet only 23% have reached the scaling or transforming stages, and another 29% are still piloting (Cambridge Centre for Alternative Finance, 2026).
The strategy is approved, the pilot works and the budget is set, but agentic AI in banking operations stays an experiment instead of doing real work at scale.
That gap between plan and practice is where the ROI goes missing, and it won’t close by itself. Every month a pilot stays a pilot, the cost keeps running and the value stays unproven.
On 15 October 2026, Nikki Constantine sat down for a fireside chat with Mambu CEO Fernando Zandona and VP of Product Adrian Congiu, who has spent nearly two decades in banking operations and fintech, and asked what separates the banks that move from the ones that stay stuck.
This article pulls out the best of their advice, in their own words, and lays it out as a path you can follow: why waiting is the riskier option, where to start, what it looks like in practice, and what to put around it.
For a long time, a bank could be slow and still hold its own, because the market around it was slow too. Fernando Zandona’s argument is that this has stopped being true.
“That has changed. It has changed completely... It starts to become more clear who is slow and who is not, and who is starting to move fast. That gap will start to increase. So as a financial institution, you need to catch up as soon as possible.”
Fernando Zandona, CEO, Mambu
What drives the change, in Fernando’s telling, is who is arriving. Newcomers come with a different mindset from most established institutions, so the distance between them and the banks that hesitate keeps growing.
The Cambridge survey shows what that distance looks like: fintechs lead traditional institutions in advanced AI adoption, 47% to 30%, and 56% of fintechs report increased profitability from AI against 34% of traditional institutions. Speed, in other words, is no longer only an efficiency measure. It is an integral part of strategy.
Fernando backs this up from his own desk. He uses an agent on Mambu’s finance data, and the example is deliberately ordinary. A bridge report, which explains what moved a number from one figure to another, used to mean a request to the finance team and a wait of a week. Add a follow-up question and it stretched to a month before he had what he needed.
Now he asks his agent, and he can ask it to “simulate these different outcomes” too. What he can do at his fingertips today, he says, is “unheard of”.
For a bank leader, the lesson is less about the report and more about the decision. A month’s wait changes what you ask for and when you act.
That pace is why Fernando keeps one line close: AI is the worst it is ever going to be today, it will only get better. The practical reading is to judge a use case by where the technology is heading, not only by where it stands today.
So waiting is not neutral. But starting is no guarantee of progress either, which is why so many banks that do begin still stall.
The conversation for most institutions has moved on from whether to try agentic AI. The harder question is what changes once a pilot works, and Fernando and Adrian each point to a different part of the answer. Fernando has a simple test for the banks that sit somewhere in the middle.
“We have high AI adoption internally. But is that translating to actual ROI? Is that really becoming a transformation, or are we just using AI to help us write our emails?”
Fernando Zandona
Adoption counts who has access to a tool. Return counts what changed in the way the bank runs, and a bank can score well on the first while stuck on the second.
The Cambridge survey points the same way. Of the financial services firms surveyed, 55% say it is difficult to measure the value of AI deployment, rising to 76% among large institutions, and only 14% see AI as transformational to their strategy.
Adrian Congiu sees the same pattern. Mambu’s customers are all thinking about how to put this technology to work and get a return, and across the market he sees a gap between ambition and what actually gets deployed: too often a proof of concept with a very limited scope that doesn’t live up to the promise.
He names two walls that financial institutions keep hitting. The first is data.
“Data is fragmented. It’s not good quality.”
Adrian Congiu, VP of Product, Mambu
An agent is only as useful as the information it can reach. Adrian puts the consequence plainly: it gets harder for an agent to reason against fragmented data in a safe, reliable and consistent way.
The Cambridge survey agrees: data availability and quality is the leading obstacle to AI adoption, cited by 40% of the financial services firms surveyed and 49% of traditional institutions.
The second wall comes from inside the bank. First-line and second-line risk and compliance teams review these platforms, and they want the right governance, auditability and permission controls in place before they will back an agent. In the usual three lines of defence model, the first line runs day-to-day operations and the second line, risk and compliance, checks that they are controlled.
Gartner’s reasons for cancelled projects start to sound familiar. It predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls (Gartner, June 2025). Unclear value is the return problem, and inadequate risk controls is the second wall.
Neither wall is resolved by raising ambition. Both become more manageable with a smaller, more contained start.
If adoption isn’t the same as return, the difference has to be owned by more than the technology team. Fernando is blunt that handing every developer an AI coding tool and declaring success is not the point.
“It’s not an engineering problem... This is governance. This is a mindset change.”
Fernando Zandona
Mambu is running the same change on itself, and Fernando says the team is learning a lot from it. His advice has three parts:
Resistance deserves attention too. “There’s a group of folks who will always love to be attached to the old ways of doing things,” Fernando says, and that is where to focus.
Once teams can build their own agents on the platform, he says, people with no coding experience start creating use cases of their own, without waiting for the whole IT organisation and with governance, compliance and security still in place.
That last clause matters: it is self-service without losing control.
The natural instinct is to start with the showpiece: the most advanced process and the most impressive demo. Adrian argues for the opposite. Start small, contained and aimed at a genuine pain point.
“Start with a very contained use case, start small, build trust, get that clear ROI use case sorted, and then expand from that into more complex things.”
Adrian Congiu
Read-only access is the key idea in that sequence. The agent can look things up, summarise and recommend, but it can’t change anything in the core, so the bank and its risk and compliance teams can watch it work on real data before anything is at stake.
Adrian’s suggested order is read-only first and action automation later. Once that has been proven and trust earned, he says, it is a lot easier to build momentum inside the bank to scale.
The process matters as much as the order.
“Ideally, something that’s a high-frequency task that is a pebble in the shoe of your employees.”
Adrian Congiu
A good first candidate tends to have three traits. It happens often, it eats up time, and the people doing it would gladly hand it over.
Fee reversal is not the most exciting use case, Adrian admits, but repetitive tasks like it take up a lot of a team’s time. Customer service, operations and payments are all places to look.
Then there is how you build. Adrian puts it simply: “It’s configuration, not customisation. We’re not deploying from scratch.”
In plain terms, the bank sets up an agent on a platform that already works, using its own rules and policies, instead of paying for a new system to be built. Whether a vendor works this way is a fair question to ask.
Larger institutions can have inertia around changing critical systems such as the core, Adrian says. For them he suggests a ‘sidecar’ project, started alongside the main system on cleaner infrastructure, where there is more room to be flexible.
Adrian describes the prize in terms of people before technology. The idea is to make the way your best banker works the standard across the team.
“You have the opportunity now to take your best banker and multiply that across your entire workforce.”
Adrian Congiu
“Every banker becomes your most knowledgeable banker,” he adds. Much of the working day goes on hunting for information, consolidating it and interpreting it before you get in front of the customer, and hours of that, Adrian says, can become minutes and seconds.
Adrian returns to a simple example: a fee query. Without an agent, an operations associate finds the customer in the core, works through the transactions, finds the fee, works out why it was charged and makes a call.
With an agent, Adrian says, the associate hands that legwork over. The agent pulls the transaction history, reasons against it and checks the fee reversal and waiver policy the bank has loaded onto the platform, then returns a summarised recommendation for the associate to review.
Today the agent works in read-only mode, by design: it gathers the information, reasons against the bank’s policy and puts a recommendation in front of the associate, and nothing in the core changes until a person acts. That already covers the searching and summarising, and it is where a bank builds trust.
The next step is action agents, where the agent asks whether the associate agrees with the recommendation to waive the fee and, once the associate says yes, carries out the waiver. A person stays in the decision, and the agent removes the work around it.
“It’s not a super sexy use case. It’s a use case that happens, though, very often.”
Adrian Congiu
Multiplied across every query, he says, the minutes add up. The benefit reaches the customer too, according to Adrian: staff can turn requests around much more quickly and understand a customer’s situation accurately, because the agent pulls their data together and summarises it.
Now the revenue side. What changes here is the trigger. In Adrian’s example, the event itself starts the work.
“Your bankers can go home at the end of the day, but your agents can stay working.”
Adrian Congiu
A large deposit lands in a client’s account. A relationship manager wouldn’t normally spot it until they read the deposit reports the following day.
On an agentic platform, the deposit triggers an agent instead. It pulls the transaction into context, reasons about the customer’s full relationship and the bank’s products, works out the best savings product and drafts an outreach.
The relationship manager walks in at 8am to a ready-made cross-sell opportunity for a product the customer doesn’t have today, with the groundwork done before the day begins.
Adrian has a name for the shift behind both examples.
“Moving from a world where you have systems of record that store and track data to systems of action that become much more proactive.”
Adrian Congiu
A system of record answers “what happened?” A system of action also asks “what should happen next?” and acts on the answer.
That is why this differs from the automation banks already know. Robotic process automation, which copies the clicks a person makes on a screen, was “very frail”, Adrian says, and often broke when a system field changed. Agents have more flexibility because they can reason against the whole context of a customer relationship.
What excites Adrian most is scale. In his banking days, he explains, banks watched “jaws”: how fast revenue is growing compared with how fast costs are growing. When revenue grows faster, the gap between the two opens like a pair of jaws and margins widen.
“This platform, this technology breaks that equation. You can now start to really grow your business without linearly growing your cost base at the same time, because your employees become that much more powerful, that much more efficient.”
Adrian Congiu
Adrian is clear about why this part matters. This isn’t software you download and play with, he says. In banking, getting it wrong has real consequences for customers. So the controls come first, and he describes three that apply to every agent action:
“We’re not deploying black box type AI systems.”
Adrian Congiu
In a regulated bank, an answer nobody can explain is hard to defend. That is why Adrian stresses that these systems were designed with the domain context of banking and finance in mind.
Under the hood, Adrian describes the three layers of Mambu Agentic. Agents, built for areas such as collections and customer service, sit on top with their guardrails. A foundation layer provides the knowledge base, the skills and the Model Context Protocol, a standard way for agents to connect to systems and data. Adrian says it links the agents to Mambu’s APIs without a rigid integration. A trust layer handles authentication, permissions, observability and logging.
The upshot, he says, is that a bank doesn’t have to think about the runtime. It can explain to its risk and compliance teams how permissions are managed, and show them the audit trail from the get-go.
What about the people? Fernando keeps it short.
“We believe that this is a better together solution versus a replacement.”
Fernando Zandona
The principle follows from the controls. Agents take on the repetitive work, and people keep the decisions and the relationships.
The Cambridge survey suggests the industry feels the same way: loss of human oversight is among the top AI risks, cited by 55% of the financial services firms surveyed, and the industry is more concerned about it than regulators are.
In a regulated environment, Fernando says, humans still need to approve actions and still need to talk to customers. “We see this as superpowers to existing operators inside the bank.”
Adrian pictures bankers spending more time engaging customers and doing value-adding work, while administrators spend more time on policy and on overseeing outcomes. The role shifts from doing the task to setting the rules and checking the results.
Conclusion
Agentic AI in banking gets out of the pilot phase one process at a time. Pick a task your team would happily hand over, begin with read-only access, build governance in from the start and keep a human on the decisions that matter. Do that, and the second use case becomes easier than the first.
Fernando’s advice is to be pragmatic and to start now: try it out, use a sidecar project and make the transformation a very big priority, or the priority, for your organisation. He calls the stakes “existential”, and says it might be for your company too.
He closed with an offer to the viewers.
“Please let us know how we can help, not only on the commercial side... We’ve been learning a lot during this process. If you want to have a conversation, learn more about how Mambu is doing it, please reach out to myself or Adrian. This is not a commercial conversation. It’s paying forward. It’s important that the whole segment upscales. We must upscale together. We all win when that happens. So please reach out.”
Fernando Zandona
Bring us the process that costs your team the most time, and we’ll talk you through where agentic AI could start.

Fernando is a technology and engineering leader with over 20 years of experience in organisations such as Amazon and Microsoft. He leverages deep technical insight gathered in hyper-growth environments to drive Mambu’s global teams to deliver optimal results for customers.

Adrian Congiu is a product leader with nearly two decades in banking operations and fintech. He previously led lending solutions at Silicon Valley Bank and specialises in building and scaling products that help banks and lenders serve their customers more effectively.
Find quick answers to the most commonly asked questions about this topic below.