Fiserv has hired a machine to help rewrite the software that quietly runs thousands of banks. The payments and financial technology giant said on May 28 that it will deploy Devin, the autonomous AI (artificial intelligence) software engineer built by the lab Cognition, to speed up the modernization of the core platforms its bank and credit union clients depend on every day.
That pitch leans almost entirely on speed. Yet the projects that replace or rebuild a bank’s central ledger rarely stall on how fast engineers can type. They stall on data migrations gone wrong, on regulators asking hard questions, and on risk committees that refuse to sign off until every edge case has been tested twice.
What the Devin Deal Puts Inside Bank Code
Cognition is the AI agent lab behind Devin, which it released in 2024 and markets as the first autonomous software engineer. Rather than autocomplete a line at a time, the agent plans, writes, tests, iterates, and ships production code on its own, working inside an existing codebase and using the same tools human developers use.
Fiserv plans to point that capability at its heaviest work. The company will use the agent for core platform modernization and other complex engineering initiatives, running tasks in parallel across large codebases to expand capacity without adding headcount. Fiserv operates some of the most widely used core account processing platforms in the country, including Premier, DNA, and the cloud-native Finxact, and a large slice of American banking sits on a Fiserv core system for banks.
Dhivya Suryadevara, co-president of Fiserv and a former chief financial officer at General Motors and Stripe, framed the move as a race against the clock.
Speed matters more than ever in banking, and our clients are counting on us to deliver. With Devin, we can accelerate modernization of the platforms our clients run their business on, ship new capabilities faster, and free our teams to focus on the work that matters most.
Why Speed Was Never the Core Modernization Bottleneck
Here is the uncomfortable part of the story. Core modernization has a long record of expensive failure, and the speed of writing code is almost never the reason.
The Years Disappear Into Testing
Replacing a bank’s core typically takes three to seven years and costs anywhere from $100 million to $2 billion, according to industry estimates. When Commonwealth Bank of Australia swapped out its legacy platform, the job ran five years. Most of that calendar goes not to authorship but to migrating decades of account data without losing a cent, then running the old and new systems in parallel for 12 to 24 months to prove nothing broke. In other words, the bottleneck was never the keyboard.
Why Boards Slow the Work Down
The other delay is human. McKinsey research on core technology economics has found that roughly 70% of large transformation efforts fail to hit their goals, and the most common causes trace back to governance, sequencing and alignment, not tooling. A faster code generator does not change what a board, an auditor, or a federal examiner will permit anyone to touch in a live banking system.
| Dimension | Traditional core overhaul | With an autonomous agent |
|---|---|---|
| Typical timeline | Three to seven years | Build phase faster, total timeline largely unchanged |
| Biggest cost driver | Data migration and parallel-run testing | Unchanged |
| Main source of delay | Regulatory sign-off and risk review | Unchanged |
| Where the AI helps | Limited | Faster code, not faster approval |
The Numbers Behind Fiserv’s Wager
Fiserv is making this bet from a position of scale. The company reported $21.2 billion in revenue for 2025, and Cognition is one of the hottest names in enterprise AI, having closed a funding round of more than $1 billion at a valuation near $26 billion the day before the partnership went public. The two figures explain why the announcement landed with so much attention.
- $21.2 billion – Fiserv’s 2025 revenue, the engine the bet rides on
- 13.86% – share of benchmark coding tasks Devin resolved end to end when Cognition first published its scores
- 42% – the proportion of US banks running on a Fiserv core platform such as Premier, DNA, or Cleartouch
The scale cuts both ways. Because Fiserv infrastructure powers such a large share of American depository institutions, faster releases could push security fixes and new features to banks more quickly than rivals can. The same reach means a single flawed change touches an unusually wide field of institutions at once.
The Governance Gap Regulators Will Probe
Fiserv is not pretending the risk away. Alongside the rollout, the company said it is strengthening governance and security controls specifically for AI-assisted development, language that signals it expects scrutiny.
It has reason to. Multiple studies over the past year found that a meaningful share of AI-written code ships with security flaws, and one analysis of pull requests generated by coding agents reported that 87% contained at least one vulnerability, with SQL injection and weak authentication among the most common. Code that compiles and passes a quick check can still hide an exploit.
Then there is the supervisory question. Bank examiners lean on model risk management frameworks such as the Federal Reserve’s SR 11-7 guidance on model risk, written for predictive statistical models, not for agents that autonomously commit production code into payment and ledger systems. The rulebook for this kind of deployment has not been written.
- Accountability: who owns the defect when an agent, not a named engineer, wrote the code
- Security: independent testing shows AI output carries injection and authentication flaws at high rates
- Examiner scrutiny: existing model risk rules were built for forecasting tools, not coders
- Systemic reach: one bug in a shared core can hit many banks simultaneously
Where Devin Already Ships Code
Fiserv is far from the first large enterprise to put the agent to work, which gives clients a track record to study rather than a leap into the dark.
- Goldman Sachs – investment banking, piloting the agent alongside its developer corps
- Ramp – corporate finance and spend management
- Zillow – real estate technology
- Lowe’s – retail
The Goldman case is the most instructive for a bank audience. The firm began piloting the autonomous coder alongside roughly 12,000 human developers, with engineers supervising the agent and steering it toward grind work such as updating older code to newer languages, exactly the kind of task that eats modernization budgets.
Russell Kaplan, co-founder and president of Cognition, called Fiserv “exactly the kind of organization where Devin creates compounding value,” pointing to its scale and the breadth of what it has to build and maintain. The caveat sits inside that praise: scale magnifies value when the code is right, and magnifies damage when it is not. Independent reviewers have flagged reliability gaps, with one closely read test completing only 3 of 20 assigned tasks and noting the agent did not reliably warn when it was uncertain.
What the Risk Committee Decides
If the agent’s output clears the risk committees and the examiners as cleanly as it clears a test suite, Fiserv will have found a genuine edge in a business where shipping a single feature can take quarters.
If the first AI-written change to a live core triggers an outage or an awkward question from a regulator, the speed story gets repriced overnight, and every bank on a Fiserv platform will be watching which way it goes.
