Post
The Future of Finance Is Autonomous
Every few decades, a wedge of technology forces its way into the calcified layers of the global financial system.
In the 1970s, it was mainframes—rigid but durable, encoding banking logic that still haunts systems today. In the 1990s, relational databases and Excel standardized information flows, enabling a wave of automation and quant strategies. In the 2010s, cloud infrastructure broke the physical constraints of data centers, unlocking distributed finance at global scale.
Now, in the 2020s, artificial intelligence is not just another wedge—it’s a chisel for legacy itself. LLMs and agentic systems aren’t productivity tools; they’re a new substrate for rewilding financial infrastructure. This isn’t optimization—it’s a generational opportunity to replace brittle systems, reprogram institutional memory, and reimagine how capital is moved, priced, insured, and governed.
Automation in Coding: The Next Trillion Dollar Asset Class
The defining asset classes of the last 50 years—cloud computing, the internet, enterprise SaaS—were all built on a single premise: labor leverage through abstraction. Each wave reduced operational friction and amplified what small teams could achieve.
Now, we’re entering a fourth epoch: autonomous software engineering.
But this isn’t just developer productivity 2.0. This is codebases that evolve themselves, infrastructure that reconfigures on demand, and agentic systems that test, debug, and deploy without human prompting. It reframes software not as a craft, but as a living system—self-maintaining, self-improving, and radically composable.
Much like cloud abstracted physical servers and APIs abstracted business logic, autonomous engineering now abstracts engineering itself. The result? A new investable layer of automation—software that no longer needs to be written to be built.
Case Study: Cognition and Devin 2.0 Launch
No company captures the future of work more vividly than Cognition AI—the creators of Devin, the world’s first fully autonomous AI software engineer.
Devin stunned the market at launch with its ability to autonomously complete GitHub issues, write integration tests, set up development environments, and deploy production-ready code—all for $500/month. Now, with the release of Devin 2.0 (read about it here), Cognition has raised the bar even higher. The model is faster, smarter, and more capable—and pricing has been radically democratized. Access now starts at just $20 initial minimum, with a flexible pay-as-you-go option for startups and small teams, and a robust enterprise offering for institutions seeking to harness Devin at scale.
Devin isn’t a co-pilot. It’s an asynchronous workforce and the future of engineering.
What sets Devin 2.0 apart:
- Cloud-native autonomy: Multiple Devins can be spun up in parallel, each tackling a discrete task within an isolated, ephemeral dev environment.
- Interactive planning with Devin Search: Developers collaborate with agents to scope work, traverse complex codebases, and trace dependencies with context-aware reasoning.
- Devin Wiki: Every project is self-documenting—automatically indexed for reusability, onboarding, and auditability.
- Multimodal orchestration: Devin synthesizes across logs, tickets, code, and documentation, enabling full-cycle resolution from root cause analysis to verified fix.
From Tasks to Transformation
Every revolution in enterprise software begins with a single task. In finance, that task has always been automation.
COBOL, a language designed in the 1950s—decades before the internet—was born of an ambitious goal: to create a universal, English-like programming language for business applications. It wasn’t elegant or fast, but it was remarkably dependable. That dependability made it the backbone of mission-critical systems across finance and government, where it still quietly powers the modern economy. Today, COBOL processes more than $20 trillion in transactions annually, with over 800 billion lines of code still in active use. Yet beneath the surface of contemporary automation lies a brittle foundation of aging infrastructure. Entire industries remain tethered to legacy codebases never designed to scale in a cloud-native world. Rather than reallocating resources toward modern, agile platforms, organizations remain entangled in the upkeep of outdated systems—trading long-term innovation for the illusion of short-term continuity.
And the cracks are showing. As Fast Company recently reported, COBOL software is so fragile that adding a new policy conflicting with existing logic can crash entire systems. Many legacy environments lack automated testing routines altogether, forcing engineers to manually write and execute tests. Even more daunting, significant portions of this code are undocumented—leaving newer engineers, including those in federal modernization efforts like DOGE, with little idea what the original logic was meant to accomplish. What remains isn’t just legacy code—it’s a form of institutional memory decay.
Meanwhile, modernization stalls. Nearly 80% of the U.S. federal government’s IT budget goes toward simply keeping systems online. Across industries, the pattern holds: maintaining what exists consumes the lion’s share of resources. Forbes reports that more than 60% of corporations still run customer-facing applications on legacy systems. What’s framed as operational stability is, in truth, technical debt—quietly compounding in the background while competitors move faster.
The last major modernization wave—Y2K—was driven by fear. Today’s transformation is driven by possibility. Legacy ETL pipelines, batch jobs, exception handling, document ingestion—these are the billion-dollar bottlenecks where agentic AI thrives.
This is not about chatbots.
This is not about email summarization.
This is infrastructure automation at trillion-dollar scale.
For Financial Institutions, The Impact is Transformational
Imagine a mid-sized bank stitched together by brittle systems—loan origination tools, risk engines, treasury portals—all dependent on aging COBOL scripts and human memory. Devin-class agents can autonomously refactor, annotate, and sustain this software stack. What once took quarters now takes days. What once required teams can now be done in parallel—without losing context or continuity.
This isn’t augmentation. It’s replacement. And it redefines the cost, speed, and resilience of enterprise software.
Zooming Out: Why This Matters for Financial Systems
Yes, software ate the world. But now AI is eating the software.
In financial services—where complexity is codified in sprawling backends, regulatory constraints, and decades of legacy code—agentic systems like Devin represent more than a productivity boost. They mark a structural shift.
This isn’t about helping developers move faster. It’s about enabling software to ship itself—and elevating engineers from line-by-line implementers to strategic system designers.
For banks and insurers burdened by brittle infrastructure and siloed tooling, the implications are sweeping: compressed release cycles, lower compliance risk, reduced maintenance overhead, and a path to rebuild institutional memory in code. In a sector where time-to-change has long been measured in quarters, Devin-class systems collapse that timeline to days—fundamentally rewiring what it means to build in financial services.
This Market Is Measurable—and Massive
Consider the numbers:
- $700B: Global financial services IT spend annually
- ~70%: Estimated portion spent on maintaining legacy systems
- <1M: Remaining COBOL programmers worldwide
- $42B → $400B: Forecasted growth of AI in financial services by 2030
Firms like JPMorgan employ 50,000+ engineers—more than many tech companies. Goldman Sachs produces 10,000+ research memos each year. How many of those could be drafted, translated, and decision-routed in seconds using AI?
Agentic AI systems don’t take vacations. They don’t lose context. They don’t need training cycles. They scale linearly with compute—and that’s the beginning of exponential financial leverage.
The Thematic Supercycle: Geography, Agents, and the Financial OS
Let’s break down where the next $100B of enterprise value will be created:
1. Modernization of Task-Oriented Systems
- The modernization of task-oriented systems demands more than migration—it requires rethinking decades of brittle software that stifles adaptability and scale.
- Aging COBOL infrastructure and fragmented ETL workflows are prime targets for transformation, unlocking trapped productivity and reducing operational fragility.
- Structured, contextual, and well-labeled data pipelines are the hidden infrastructure that determines whether AI can deliver precision or just noise.
- Startups embedding deeply in compliance, risk, and ops are not just building tools—they’re reconstructing the internal logic of institutions, replacing tribal knowledge with scalable systems.
2. Geographic Asymmetry Creates Arbitrage
- Emerging markets without legacy constraints are skipping straight to modern architectures, creating structural advantages over slower-moving incumbents.
- Nubank exemplifies what’s possible when you design a bank around real-time decisioning, automated fraud detection, and machine learning at the core—not bolted on.
- Many European banks are still stuck with batch-based processing and outdated compliance infrastructure, despite regulatory nudges like PSD2.
- UPI is more than a payments layer—it’s a modular, API-first system that’s uniquely positioned to support AI-enabled financial services at scale.
3. Agentic Systems in High-Complexity Workflows
- Asset Management: LLMs synthesize macro data, earnings calls, and alt-data into research instantly.
- Insurance: AI chains adjudicate claims, detect fraud, and optimize pricing at scale.
- Banking: Credit officers and RMs use AI copilots to surface customer insights and risk indicators.
- Compliance: Trade surveillance and regulatory parsing are already LLM-native and real-time.
Earnings, Efficiency, and the Code Dividend
Why does this matter?
Because in finance, efficiency is margin. And margin is market cap.
Ken Thompson once said, “One of my most productive days was throwing away 1,000 lines of code.” In the AI era, value doesn’t accrue to more software—it accrues to less manual intervention.
A recent McKinsey study estimates that banks deploying intelligent automation at scale can reduce costs by 20–35%, unlocking hundreds of billions in enterprise value. For an industry where even 100bps of ROE matters, that’s not just upside—it’s existential.
The Operating System for Money Is Being Rewritten
AI won’t replace finance professionals. But institutions that deploy AI-first systems will replace those that don’t.
For over a decade at Conversion Capital, we’ve partnered with founders transforming the infrastructure of finance, applying AI to real-world problems, and building the operating systems that underpin today’s economy. Our experience gives us a clear view into where inefficiencies persist, how opportunities compound over time, and what it takes to scale intelligent systems through disciplined go-to-market execution in highly regulated markets.
This isn’t a thesis. It’s a blueprint. And it’s only just beginning.