Revolut PRAGMA: The Foundation Model for Money

In this episode of the Macro AI Podcast, Gary Sloper and Scott Bryan unpack Revolut PRAGMA, one of the clearest signals yet of where fintech and AI-native banking are headed. 

PRAGMA is not a chatbot or a simple banking app feature. It is better understood as Revolut’s financial intelligence layer — a foundation model designed to understand customer behavior, banking events, risk patterns, product engagement, and how people actually move money. Gary and Scott explain how PRAGMA differs from AIR, Revolut’s customer-facing AI assistant, and why the real story is not just conversational banking, but the deeper intelligence engine underneath it. 

The discussion breaks down how PRAGMA treats financial activity as a sequence of events: salary deposits, card transactions, currency exchanges, subscription payments, stock trades, product clicks, and fraud signals. When organized over time, these events become something like a financial language that can help support fraud detection, credit scoring, product recommendations, customer engagement, and more. 

Gary and Scott also explore why this matters for business leaders beyond fintech. PRAGMA shows that AI advantage is shifting from generic tools to proprietary intelligence built on domain-specific data. Revolut’s model highlights the power of usable data, shared AI infrastructure, agentic user experiences, and governance. 

The episode also covers PRAGMA’s limitations, including why anti-money laundering often requires graph intelligence rather than only customer event histories. The broader takeaway: AI-native finance will likely combine sequence models, graph models, language models, anomaly detection, rules engines, and human review. 

For banks, fintechs, and enterprise leaders, the message is clear: AI is moving from feature to infrastructure. The future competitive advantage may not be the app, card, branch, or product menu — it may be the intelligence layer that understands every customer, every event, every risk signal, and every opportunity in real time. 

About your AI Guides

Gary Sloper

https://www.linkedin.com/in/gsloper/

Scott Bryan

https://www.linkedin.com/in/scottjbryan/

Macro AI Website: 

https://www.macroaipodcast.com/

Macro AI LinkedIn Page:  

https://www.linkedin.com/company/macro-ai-podcast/

Gary's Free AI Readiness Assessment:

https://macronetservices.com/events/the-comprehensive-guide-to-ai-readiness

Scott's Content & Blog

https://www.macronomics.ai/blog

Transcript

Welcome back to the Macro AI podcast. I'm Gary Sloper.  And I'm Scott Bryan. And today we're talking about Revolut Pragma, which we think is a good signal about where FinTech is heading and a topic that we've received a few questions about over the past few weeks.  And it seems like we have a lot of FinTech listeners, which I think is really interesting and we appreciate that.

01:24
ah And that also includes college students, which is great to see because we know that there's so much pressure on college students coming out of school with AI.  Hopefully you're finding this show helpful.  But before we get into Pragma, it is worth setting the  context on Revolut. Revolut is a London founded fintech that has grown from a mobile first foreign exchange and payment app, really into one of the world's largest neo banks and financial super apps globally.

01:53
It now offers a broad set of services across personal banking, international money movement, investing, crypto, subscriptions, and business accounts. That broad product surface is important because Pragma is not being built around one narrow banking product. It's being built around a huge stream of customer financial behavior across the Revolut ecosystem. Yeah, Gary, that's a good overview. uh Pragma from Revolut is not just another

02:20
AI chat bot or simply a uh feature that's inside of a banking app. I think it's better understood as Revolute's  financial intelligence layer or  a foundation model designed to understand customer behavior, understand banking events, risk patterns, product engagement,  how people actually move their money. Yeah. And Scott, think that's a good point. And it matters because the next competitive advantage in FinTech may not just be the app or the card or the product menu.

02:50
menu or  the banking license, it's probably the intelligence layer underneath all of that from a financial perspective. Exactly. Yeah, totally agree.  And I think the  first thing to clarify is the difference between Bragma and then the customer facing assistant that's inside the app called AIR.  And so AIR is AI by Revolut or AIR.  And like I said, it's the customer facing assistant that's inside the app.

03:18
So it helps users ask questions, uh understand their spending, manage parts of their financial life, really, and interact with their account in a conversational manner. So, so Pragma is, is deeper in the stack. Pragma is actually the model family that helps Revolut understand all that financial behavior at scale.  It's trained on banking event histories. So transactions, app activity, trading activity.

03:46
customer communications, product use,  and all of these things over time. Right. And so AIR is the assistant, Pragma is part of the intelligence engine. So we just wanted to kind of split that out for distinction. And I think that distinction is important because most people hear quote unquote AI in banking and think of the support bot, the chat bot that you've probably interacted with. ah

04:12
But the bigger shift is that financial institutions are starting to build models that understand the behavior of money itself. And I think the phrase,  the behavior of money is really key here. A salary deposit is an event. A card transaction is an event. uh A currency exchange is an event. A subscription payment is an event. A stock trade is an event. A product click is an event. A fraud signal is an event. But when those events are organized over

04:41
time over that period, they become something like a financial language for the environment. Yeah. Perfect. And I think our listeners understand that it's not language in the traditional sense.  Money data is, is easily structured.  a Trent transactions, seven amount, you know, currency merchant, uh, their categories, their timestamps, there's even customer context built in there, et cetera. So

05:08
That's why generic LLM is not always the best architecture. A general model  may understand words like  salary or credit card, but, but pragma is designed to learn the actual pattern of financial behavior and even down to the customer level. And I think that's kind of the broader enterprise lesson for, for this  episode is that the most valuable AI systems may not be generic chatbots.

05:37
they're probably going to be domain specific models built around the structure of any particular business. eh that's a point. And, and so if you're listening today, you're probably asking, okay, what does pragma actually do? And its main job is think of it this way to,  create deep mathematical representations of financial behavior at,  at, the basis. So those representations can then, you know, support things like we were just talking about fraud detection,  maybe credit scoring product recommendations.

06:06
communication engagement, lifetime value prediction of that particular client,  recurring transaction detections, and better context for the conversational banking for you as a client. ah Instead of building a separate model for every task,  So every, know, trying to build something for each one of those that I just mentioned. So Revolut is building a shared foundation model that can be adapted across many tasks that are required in that particular banking environment. Right. Yeah. And I think that's

06:35
kind of the economic power in that model is reuse. So a traditional bank might have a model for fraud, another for churn, another for credit risk, one for marketing, customer recommendation, product recommendations and things like that. And for customer service, obviously, we've talked about that a number of times in customer experience, but Pragma points toward a different operating model.

07:02
you know, train one shared backbone on massive financial event histories, and then adapt it across lots of workflows. And the scale is important there because, know, the available technical details inside that pragma was trained on, I think it was 26 million users, 20 plus billion events, 200 plus billion tokens across 111 countries over two years. So that gives Revolut a really powerful

07:31
proprietary data asset. Yeah, that's huge for them, especially, you know, not just in the short term, but the long term makes it extremely sticky and valuable. Yeah. And there are Neo  there, like we talked about there and there are Neo banks. there it was easier for them to do from the ground up versus, know, some of those big legacy massive banks. Yeah. Yeah. You bringing a lot of tech debt with you, you know, in those types of environments.  Um, and I, and I think if we explain the architecture and kind of

07:59
simplistic terms. Pragma has really three major parts. There's the profile state encoder that understands account context. There is an event encoder that reads individual events like, you know, we were talking about before, or app actions,  trades or communications. And then there's a history encoder that looks across the full sequence and creates a contextualized representation of the customer's financial behavior. So it's not just asking, is this transaction unusual?

08:26
It is asking, is this transaction unusual for this customer at this moment, given the customer's recent pattern as a client within this institution? So it's, very interesting and it's doing it at warp speed versus a human looking at this. Yeah, exactly. And I think that that customer context matters. So for example, a $500 transaction means something different right after payday than it does the day before rent is due.

08:56
And  a foreign transaction means something different for a frequent traveler than for someone who's never even made an international purchase or  travel. And that's why, uh, pragma is,  I think that's exactly what it's, it's trying to learn, not just the event, but the event in the context  and in the context of that consumer, that user. Right. Right. And, um, you know, Revolut is  really

09:23
meaningful improvements across credit scoring  and other areas such as fraud detection and  product recommendation for that banking client. ah We talked about communication engagement and other prediction tasks. So we should be careful not to overstate it. There are, you know, revolutes disclose benchmark results, not a third party audit of every production use case, but the direction is important. The model seems strong as where the signal is spread across many events over time.

09:52
And hopefully as it continues to learn, it'll just become that much more beneficial to the end user. Yeah. Yeah. And that's, that's why this matters. think financial behavior is, is very rarely explained by one variable. Um, so  risk engagement product, but in those things that we talked about  and fraud are usually hidden across lots of different signals and pragma can learn patterns that are difficult for traditional feature engineering to capture.

10:20
So product recommendation, for example, becomes more than, uh you know, this customer belongs to this segment. It becomes, given this customer's recent behavior and financial rhythm, what is actually relevant to them right now. Right. Yeah, that's a point. And I think that actually really brings us to air and the user experience. Banking apps have historically been menu based, especially if you've been banking for a long time. You're familiar with that. You log in.

10:47
Click cards, click accounts, click settings, click transfers, you know, other things like clicking on budgeting and support. It's all click, click, click, click, ah Air points towards a different model. Instead of learning the bank's menu structure, the customer expresses intent. Why did I spend so much this month? Which subscription am I paying for? ah Can I afford this trip? know, freeze my card. What changed my spending? And this sounds like a UI change, but it's actually much bigger. And in

11:14
We've probably seen some smaller examples. know, Zelle is out there and kind of does a little bit of that ah interaction, you know, with, with the intent, but this is on a much larger scale ah to be much more scalable. Yeah. You're definitely starting to see it across these apps and the financial institutions have a lot of  talent and experience and they're, they're getting there. uh But this, think this is the very beginning of the agentic user experience in banking. Yeah. And point. Yep. And that's going to be relevant to,

11:43
all kinds of other industries like we mentioned. So menus aren't going to completely disappear.  Banking still needs confirmations, permissions,  audit trails and things. But the main primary interaction can shift from navigation to actual intent.  think back,  those generic chatbots can define things like budgeting, ah but a money intelligent...

12:10
assistance, intelligence assistant can actually explain your budget. can tell you why your spending changed, um what bills are coming exactly when, and when cashflow could get tight for you. And that's the difference between financial information  and financial intelligence. Yeah. And over time it can become proactive. Your bank could warn you about cashflow pressure, identify  unused subscriptions.

12:38
suggest moving out of cash into a higher yield account, example, ah flag suspicious activity, or even just help with the, you know, optimizing your spending. And, that is very useful, but it's also very sensitive. And I think, I personally think we'll probably see, you know, some of the demographics, you know,  embrace it, maybe not so much.  You may have an older generation that says, Hey, I want my menu. I, know, I'm very structured and I want to do very specific things in a younger generation that is growing up with.

13:07
you know, generator of AI now may say, Hey, I just want to be able to speak  and the system knows exactly what I want to do. Yeah. It's yeah, definitely sensitive. think the younger guys uh and girls are definitely less sensitive, but a model that understands your spending better than you do, uh, can definitely be helpful, but it can also feel pretty invasive.  And  I think money naturally reveals a lot about people, know, their lifestyle.

13:34
financial stress, family needs,  typical travel, health related purchases, uh your typical annual donations  and other pressure points. So trust becomes the product boundary. And I think the best AI bank will not just be the one that predicts the most, it'll be the  one that earns the right to use prediction on behalf of the customer. So, you know, that really strong trust level.

14:02
Yeah, and to that point, I think it's one of our topics we've mentioned many times on this show. And that's where governance matters here. If pragma style models  influence credit decisions or fraud detection or product eligibility  or even  financial recommendations for you, the regulators will know how those decisions are made and will want to know that. ah In credit especially,  think of it this way, ah explainability is not optional.

14:32
If a customer has denied a product or offered  less than favorable terms, the institution needs a clear explanation here. So artificial intelligence in finance is not just about model performance, it's about  things like audibility, fairness, privacy, security, human oversight, and really operational resilience because you will need to explain that. Why did I... uh

14:59
not receive  the lowest interest rate available. Why am I at a five or six X interest rate? Or why was I just flat out denied this  loan, for example,  and being able to back that up from a governance standpoint?  Yeah. Yep. 100%. Yeah. Explainability. um And just to note, Pragma also has limits. uh One of the most interesting disclosures is that Pragma did not...

15:26
outperform Revolute's existing AML or anti-money laundering systems. And that makes sense, I think, because it's going to take a lot of work because  anti-money laundering is not just a sequence problem, it's a  greater  network problem.  money laundering often requires  understanding relationships between accounts,  counterparties, merchants, devices, ah geographies,  all kinds of things. Payment flows.

15:55
Pragma is designed to understand a customer's event history  and anti-money laundering often requires uh a graph type of intelligence.  So the lesson is not that, know, Pragma has a weakness. It's that the lesson is that different financial problems  require an actual different AI architecture. Yeah. Good point. And just think with all of the data and history over time, it may be able to  detect

16:23
um, anti money laundering much more faster or, or predict potential scenarios or even identify money laundering that was overlooked. You know, once,  once the architecture is built and the training data has been there, it could actually potentially, you know, has  the potential to improve.  I'm sure, and I'm sure they're working on those graph architectures to build. Yeah.  Yeah.  Yeah. And it just, you know, the, the, the time to.

16:51
resolve those  money laundering scenarios would be  interesting.  And to your point, Praga is not a magical banking brain. It's a powerful model ah for financial event histories, but AML reminds us  that,  I think of it this way, financial crime is relational. So a customer timeline is not the same as a criminal network. And that's something to keep in mind as we're talking here today.

17:19
Yeah, just to kind of reiterate that I think the future financial AI stack will probably combine sequence models, graph models, language models, anomaly detection,  rules engines, and then the human in the loop, the human overview component. And I think that's the  real architecture of AI native  finance and it's happening now. Yeah, that's a point. We should probably also talk uh about competitive implications. So Revolut is

17:48
interesting because it has a wide product service and high frequency digital engagement. So every payment transfer, foreign exchange, FX, you know, conversation, communication subscription, uh, what else? Uh, investment action or actions and, product click becomes more of the data flywheel. We've talked about the data flywheel on, on our last episode. Um, the more useful the app becomes, the more customers use it. So the money that

18:16
they  use uh it will be  more data that Revolut has. So the more transactions, the more ah intent, it's continuing to collect that data. So ah that only improves the model. So the model becomes more efficient and more beneficial, not just for the user, but hopefully for all users. And the better the model becomes, the more useful the app can become holistically and kind of set that benchmark and help with things like governance and

18:45
and fraud and other areas. Yeah. That's the AI flywheel. And a lot of institutions are just, just getting that, getting that moving. Um,  and the moat is not just data. Lots of banks have data. The moat is usable data plus architecture, plus infrastructure, plus product integration, and then governance, we talked about. So traditional banks  may have enormous historical data, but it's often fragmented across core systems.

19:13
card platforms, lending systems, fraud tools, CRM, digital channels, revolutes advantages that it's business is digital, higher frequency and it's increasingly AI native. Yeah. So for a traditional bank, the answer is not necessarily, you know, build pragma tomorrow. The answer is prepare for this future, build unified customer event histories, modernized data pipelines, um, you know, improve identity resolution, create a model governance.

19:44
uh for this for this entity that you're supporting today So start with a lower risk use case then you can carefully migrate into things like fraud and credit and personalization So you don't have to completely move the ship You know to a 90 degree or 180 tomorrow, but you need to prepare that this is where what's coming down?  The pipeline. Yeah, that's why we thought that uh this Revolut program would be just a good overview and

20:11
business leaders outside of FinTech can pay attention to this too, because every, every industry has event histories. ah Think about manufacturers, they have machine events.  And uh we just talked about physical AI on a recent episode. Retailers have customer journeys. ah Healthcare has patient journeys. Telecom, we love telecom. They have network events. ah Logistics has shipment events.  The question is whether

20:39
those events are organized on a foundation for intelligence. And I think that's kind of the universal lesson for this, for this episode. Yeah. And you know, there's also a sovereign AI angle here. should probably mention as well. Uh, yeah. Yep. That's a good one to bring up. Yeah. So, so Revolut appears to be building more of its own foundational technology and using things like the European AI infrastructure through Nebius.  Um, That matters in finance because institutions care about

21:09
things like data residency, uh operational resilience, vendor lock-in, audibility, and regulatory control.  So making sure that that information and everything that's transacted stays within that country and is, is, you know, locked down pretty hard. Yeah. Yeah. Sovereign AI definitely a big topic and something that anybody that's delving into this needs to be aware of. not every company should build its own foundation model.

21:39
but the largest financial platforms may decide that the intelligence layer that understands customers and risks is too strategic to completely outsource.  that's where we go back to that build versus buy question. And this time, obviously, it applies to AI native finance. So some capabilities can be bought and some can be partnered, some can be built, but leadership teams really need to know which AI capabilities are strategic. And for  Revolut, the model that

22:09
understands financial behavior is obviously strategic to them. Yeah. Good point. All right. So let's bring all this home.  In the best case,  Pragma and Air point towards a bank that becomes more useful. It helps customers understand their money, reduce their waste,  fraud, optimize cash, and hopefully make better decisions. In the bad version, AI becomes a way to push more products,  increase fees.

22:35
uh more borrowing under the cover of personalization. So the trust model is everything and users will quickly fill this out. Yeah, 100%. uh So AI banking,  and I'm sure there are lots of others that will apply to this, but in  specific to AI banking, it  must be aligned with the customer. So if the assistant feels like a  trusted CFO or financial advisor, people will love it. If it feels like a sales engine disguised as uh advice,

23:04
customers and regulators both will push back.  And that's why the future of banking AI will be both uh technical and then you've got the major obviously ethical considerations. Yeah, yeah, they don't want the wolf and sheep's clothing. ah So let's summarize the big takeaways. AI is moving from feature to infrastructure. Proprietary data matters, but only if it's usable. ah

23:30
Domain specific models are going to matter. The interface is shifting from menus to intent, which we talked about governance becomes a competitive differentiator and AI native companies can compound faster because every interaction improves the intelligence layer for their customers. Yeah. Perfect. Great summary. Um, and I just add that that pragma is not, you know, chat GPT for banking in the simplest terms. Yeah. It's a, it's a foundation model for, for money.

24:00
And I think the future bank may not be defined by, you know, branches and debit cards and mobile apps. It's going to be defined by the intelligence layer that understands each and every customer, all their transactions, their risk signal signals,  and  every opportunity for them  in real time.  Good point.  And for business leaders outside of finance,  the message is that every company has behavior patterns. Every company has event histories.

24:28
Every company has proprietary data that could become intelligence if it's organized correctly. So the question is not just what AI tool should we buy? The better question is what proprietary intelligence layer should we be building? Yeah. And  Revolut by Pragma and their AIR user interface is a great fintech story. But like  you just mentioned, the lesson is broader than that.  And AI advantage is moving from generic capability to deep

24:58
proprietary intelligence. Good point. I think that's a great place to leave it for today's episode. Thanks for listening to the Macro AI podcast. We appreciate all the subscribers, the likes, the questions that we share with your network. Until then, we'll see you later.