🧠 Legacy Finance is Perfect for AI

Welcome to Fintech Brainfood, the weekly deep dive into Fintech news, events, and analysis. You can subscribe by hitting the button below, and you can get in touch by hitting reply to the email (or subscribing then replying)

Weekly Rant 📣

🧠 Legacy Finance is Perfect for AI

Nearly all money moves as structured text files. AI agents communicate in markdown. Think about that for a second.

Hear me out.

What if not upgrading finance from COBOL and flat files was a superpower all along? SWIFT messages, ISO 8583, NACHA files, FIX protocol—they’re all structured text with specific field positions and delimiters. They’re structured markdown.

And guess what.

AI is all about using structured markdown (.md files) to build agents who can work together, share knowledge, and tackle more complex tasks than fit inside a context window.

So here’s something that hit me like a truck.

What if the financial system accidentally built itself in the one format AI is perfect for?

The banks didn’t fully modernize. They didn’t fully move to APIs. They didn’t fully adopt microservices. Underneath it all, at the very bottom of the stack is an on-prem mainframe and text files. And that stubbornness accidentally positioned them for the AI era better than anyone who “did it right.”

It’s all still there, begging for AI to use it better.

Digital transformation didn’t replace the mainframe.

When you see a mobile app from your bank, you’re looking at a pretty screen that speaks to some internal digital systems that ultimately pull their data from the mainframe. And the mainframe is written in COBOL.

Digital transformation hasn’t made a dent.

COBOL processes $3 trillion in daily commerce. It runs 95% of ATM transactions and powers 43% of U.S. core banking systems. The estimated cost to replace it all? $4–8 trillion.

Nobody is replacing it. Nobody was ever going to replace it.

Every decade, a fresh wave of technologists arrives to “fix” finance. In the 2010s, it was the fintechs. They look at the mainframes, the COBOL, the flat files uploaded to SFTP servers, and they see a system begging to be torn down.

The geology of finance

The infrastructure and code in financial services is like sedimentary rock. Each layer builds on top of the other:

• Channels: At the top, the branches, online and web applications you see are all pretty modern

• Digital infrastructure: Underneath those are modern microservices and digital systems

• Finance infrastructure: Underneath that are specialist fraud, AML, CRM systems

• Legacy core: Then at the core the mainframe. Running millions of lines of code.

• Rails: Finally networking and messaging layer (SWIFT, ACH, Cards, FIX).

Over the decades, code was added to the mainframe core systems to respond to every new law, regulation, and crisis—the 1997 Asian Financial Crisis, Black Wednesday, the Russian default. These cores also handle all the complex laws, edge cases and regulatory reporting like: Money Laundering Control Act of 1986, Gramm-Leach-Bliley in 1999, the PATRIOT Act after 9/11.

Each layer added complexity. Many of the original architects have long since retired. Their institutional memory is lost to the mists of time.

This all runs. It all works.

When you’re using TradFi today, you’re using a modern marvel. Code that runs around thousands of laws most people never think about.

And everyone keeps trying to replace it

Legacy Systems Create their Own Challenges

In 2020, Citibank accidentally sent $900 million instead of a small interest payment to Revlon’s lenders. The root cause wasn’t human stupidity. It was Flexcube’s user interface: the only way to process an interest-only payment required entering it as if paying off the entire loan, then using three cryptic checkboxes—FRONT, FUND, PRINCIPAL—to redirect the principal.

This is not a Citi problem. This is an industry problem.

Now imagine that overworked person replaced by AI, overseen by another AI, and another—checking and triple-checking for errors faster than you can blink. AI doesn’t just reduce errors. It makes the shit software less of a problem and easier to dismantle and displace.

Oracle clearly agrees. They launched their Agentic Platform for Banking in February 2026, shipping pre-built AI agents across the entire banking lifecycle. The very system that caused the $900 million error is now getting an AI wrapper. An agent providing a natural language confirmation—“You are about to send $900M to 315 lenders; this appears anomalous”—could have prevented the error entirely.

Most People don’t Understand Finance Infrastructure.

Financial services at a distance is infuriating and weird. With enough time in the industry, you start to learn its quirks and collect fun war stories. For example, Kristen’s story about the bank that gave customers money before they’d received the payment:

I once spent a Saturday night on the floor of an IBM data center, trying to fix a customer-facing website and the mainframe running the payments system after an outage. Into the early hours of the morning. That mainframe ran 95% of debit card volume for a European country.

Legacy mainframes, IT, and files run the world, but they’re so incredibly hard to “transform” or change. Because they take so long to learn and understand.

Digital Transformation failed because it couldn’t comprehend the entire legacy infrastructure

Banks spend literally billions on digital transformation. It’s the buzzword that has kept consultants in work for at least two decades.

The only thing that changed was what new silver bullet was designed to be the cure-all.

• Services-Oriented Architecture (SOA) promised to decompose the monolith.

• Cloud migration promised lower costs and faster development velocity.

• Enterprise CRMs promised integration.

So naturally, when AI came along promising the same thing, the seasoned banker in me sighed. Oh, sweet summer child. We’ve had false dawns before.

But this one is different. And I don’t say that lightly.

1. AI doesn’t add a layer above the mess. AI is water—it flows into every crack between the rocks. It doesn’t need the architecture to change. It works with what’s there. And it can produce new code at an astonishing rate.

2. Before AI, no single person could fit the scale of a legacy banking system into the context window of their brain. So we formed committees. Broke things into tasks. Hired McKinsey to build governance frameworks. The entire architecture of enterprise IT exists because human comprehension doesn’t scale. 

AI agents don’t have either limitation.

They work with what’s already there.

AI isn’t digital transformation. It’s an augmentation.

AI can comprehend entire codebases. They can build context by communicating with each other. They don’t form committees and use Microsoft Teams or Office. They use markdown.

• One agent runs in a loop, learns lessons, writes down how far it got, and hands it to the next. 

• They hand that to their manager agent. 

• They leave notes for each other. In markdown.

• Orchestrated by frameworks.

The financial system’s most “ancient” formats—the fixed-width files, the batch jobs, the ISO messages—are exactly the substrate AI works best with. Structured text.

All of finance is flat files. And guess what ISO messages are? Structured markdown.

I heard one Chief AI Officer say to me:

This made me think there’s a new definition for enterprise modernization.

• Digital transformation: move to cloud, adopt APIs, kill the mainframe

• AI augmentation: Keep the mainframe. Wrap it in AI. Let agents handle the translation.

Digital transformation has always been hard and expensive. 70% fail completely, with some resulting in massive customer outages and CEOs being fired. 

Maybe cloud adoption isn’t a panacea? When we saw AWS East 1 suffer an outage the large banks that hadn’t gone fully cloud native breathed a sigh of relief. Because there is still an upside to running your own infrastructure. And it’s less of an impediment to using AI.

AI can run on-prem (IBM’s Watsonx on Z, NVIDIA NIM microservices on Temenos). Banks may have been refuseniks long enough to see the world’s polar axis shift to their advantage.

On-prem is sexy again.

That said. Hiring McKinsey to build you a governance framework for legacy systems is probably not the AI-forward way of doing it. I don’t like to think in absolutes. But AI Augmentation is quickly going from pipedream of the C-suite to a practical reality. 

There’s already tooling for LLMs to understand flat files and Mainframe code.

There’s an open-source ISO 8583 simulator using LLM integration to explain raw transaction messages in plain English and generate messages from natural language prompts like “$50 refund to Mastercard at ACME Store.”

Gridworks AI’s SwiftParser handles MT103, MT202, MT515, MT700, and MT950 SWIFT messages. Turns out AI can read the smoke signals. And translate them into hundreds of languages. And more importantly, code.

• Amazon’s AWS Transform launched as what AWS calls “the first agentic AI service for modernizing mainframe workloads at scale.”

• Microsoft built a COBOL Agentic Migration Factory with Bankdata, a Danish banking consortium covering 30% of the Danish market with 70+ million lines of mainframe code.

Their early experiments with GPT-4 produced what they candidly described as “a good mix of educated guesses and hallucinational gibberish.” But with proper agent architecture and context management, output quality became “surprisingly good.”

That sounds about right. Because proper agent architecture and context management is the meta skill of the 21st century.

What this means for you

Thinking AI-first, thinking in markdown files and agent orchestration are your new must have skills.

The previous waves of digital transformation failed because they tried to build above the mess. AI works because it flows into the mess. It doesn’t need a clean architecture. It needs a context window and a corpus of notes left by other agents.

AI needs markdown, it needs flat files, and it needs access to critical systems. That is an almighty governance challenge. But my hopefulness comes from seeing just how much AI has already been adopted in customer support and the back office. I would not have expected that.

My caution comes from watching things go so. Fucking. Slow. In big orgs.

But the trend line is clear in AI, the acceleration is coming and they’re getting better, faster. Autonomous task capability doubling every 4 months. Claude Opus 4.5 now handles tasks taking human experts approximately 5.3 hours at 50% reliability. 

AI agents handling week-long autonomous tasks by 2027–2028. Orchestrated they can tackle giant code bases, flat files and the complexities of the legacy infrastructure.

AI is the biggest thing to happen to finance since COBOL.

All roads lead to the mainframe. 

But now AI speaks mainframe. 

ST.

4 Fintech Companies 💸

1. Kairos - The Multi Prediction Market Trading Platform

Kairos provides a single terminal and dashboard for traders who want to buy and sell events contracts across multiple prediction market venues. Kairos brings together real-time data, analytics, and low-latency execution in a single place. The team are ex CBOE quants and worked on ML engineering at NASA. 

🧠 You know when you see something and you go “oh that seems so obvious why did nobody do that already?” This is clearly aimed at the pro trader who operates across multiple venues, and won’t be retail-facing. Products like this will be needed if events contracts graduate from retail to Wall Street. 

2. Murfee AI - The governance layer for agentic AI and legacy systems

Murfee takes user intents and pre-screens them through its governance layer before submitting a request to an underlying legacy system. The goal is to detect any drift from human intent and measure, then audit the success of intent to outcome.

🧠 I had to read the website about 6 times and I think I understand it now. It’s a very jargon heavy page with more (™) badges than any other I’ve seen for a start-up. With all of that said, a governance layer that is performant sitting between human intent and legacy system is an idea that resonates. 

3. EnFi - AI Agents for Lending

EnFi allows lenders to screen and underwrite deals from unstructured documents, turn borrower reporting into risk assessments, and create a single source of truth for all data. The team says the platform is fully explainable and they have roles like relationship managers, underwriters and credit officers.

🧠”Agents as a service” is the new staff augmentation. It’s often cheaper and better performing too. It’s also the type of regulated workflow that relies heavily on knowing what a valid outcome looks like (evals). You have to wonder though, with the tech sell off because of the “Anthropic effect” will these companies have a moat?

4. Dost - AI Accounts Payable for mid-market SMBs (UK)

Dost helps automate invoice capture, review, and payment workflows with “100% accuracy.” It then creates a single dashboard for finance teams to see all invoices, payments and workflows. The company claims to reduce costs by 80% on average,

🧠80% cost reduction is a no-brainer. I just wonder again, how long it is before this is a feature in Claude?

Things to know 👀

1. Bridge receives a conditional OCC approval to form a National Trust Bank.

Bridge — Stripe's $1.1B stablecoin acquisition — just got conditional OCC approval for a national trust bank charter. Meaning Bridge can now operate, custody, stablecoin Issuance, reserve management, and orchestration across blockchains with direct federal supervision.

🧠 This gives their ever-larger clients more confidence in the provider of stablecoin services. Bridge already powers custom-branded stablecoins for Phantom, MetaMask, Hyperliquid, and Klarna. A federal trust charter gives all of that a regulated banking backbone. The next phase may be even larger, more regulated clients.

🧠 The conditional approval is only step one. The harder part is the 18-month organizational phase — proving they can operate a federally chartered bank through mock and live exams (which all new trust charters have to get through). Regulation is a credibility marker. But it's earned, not given.

🧠 The ABA sent a letter to the OCC last week urging them to slow approvals down 🤦 . Where the bank lobbies see a threat, I see upside. Deposits are money at rest. Stablecoins are money that moves — and settles 24/7, instantly. That's a new rail, not a replacement for deposits.

🧠 YouGov and Artemis found 77% of consumers want their bank to provide stablecoins. That’s the opportunity. The $308B stablecoin market has a legal framework now, and the regulation is coming. The banks that lean in will build new revenue lines. The ones that lobby to slow it down will watch their clients find another provider.

2. The CFTC “reaffirms” exclusive jurisdiction over prediction markets.

Per the CFTC: “The Commodity Futures Trading Commission today filed an amicus brief in the U.S. Circuit Court of Appeals for the Ninth Circuit confirming its exclusive jurisdiction over the U.S. commodity derivatives markets, including event contract markets commonly referred to as prediction markets.” Platforms like Kalshi and Polymarket are currently facing legal cases in multiple states over allegations they’re more akin to gambling than derivatives.

🧠 Some states are arguing about values. Utah’s Governor Cox (one of only two states where gambling is completely blocked) took to Twitter to say, “I don’t remember you having jurisdiction over Lebron James rebounds.”

🧠 Others about revenue. Massachusetts, Nevada, and Connecticut have either sued, taken enforcement action, or questioned the legitimacy of these contracts. These states stand to lose gaming revenue, which is collected locally.

🧠 Sports betting very clearly falls under state jurisdiction. So when you see sports contracts - you can see why there’s a kerfuffle. Technically, the instrument may be different, but the direct competition for consumers is clear as day.

🧠 Legal proceedings are rarely over quickly. This could drag on for years, and its overshadowing an important point.

🧠 Some of the data is useful. All of this is in the same week when research showed Kalshi’s FOMC forecast error is better than Fed Funds.

Good Reads 📚

1. The self improvement loop is coming

The effective workforces of AI labs will 10x and 100x over the course of 2026 as AI begins to do most of the AI improvement research. The objective of this workforce will be to make themselves smarter. The debate should not be about whether this loop will occur, but how it does and what its implications are. 

Today models are not coming up with novel ideas, but they are an army of junior researchers that can execute experiments, and very soon, those will be able to run over multiple days. And if today 800 engineers are achieving a roughly 4x efficiency improvement YoY (per Anthropic), by grinding through these experiments, than could 8000 engineers achieve much more?

Add to this that by the end of the year some of the massive Capex from 2024/2025 starts to come online and we’ll see if scaling laws are back.

Tweets of the week 🕊

That's all, folks. 👋

Remember, if you're enjoying this content, please do tell all your fintech friends to check it out and hit the subscribe button :)

Want more? I also run the Tokenized podcast and newsletter.

(1) All content and views expressed here are the authors' personal opinions and do not reflect the views of any of their employers or employees.

(2) All companies or assets mentioned by the author in which the author has a personal and/or financial interest are denoted with a *. None of the above constitutes investment advice, and you should seek independent advice before making any investment decisions.

(3) Any companies mentioned are top of mind and used for illustrative purposes only.

(4) A team of researchers has not rigorously fact-checked this. Please don't take it as gospel—strong opinions weakly held

(5) Citations may be missing, and I’ve done my best to cite, but I will always aim to update and correct the live version where possible. If I cited you and got the referencing wrong, please reach out