Chase's CIO stopped hiring engineers to write code — and won't go back

Gill Haus, chief information officer at Chase, no longer hires engineers to write code. The consumer bank, part of JPMorgan Chase & Co., now hires them to decide which code to write, an inversion Haus described as "a big difference" from practice as recently as six months ago.

The comment lands against a stubborn backdrop. 91% of professionals say their firm still falls short on AI, and 45% admit to using shadow AI tools at work. For one of the largest retail banks in the United States, the question is no longer whether to deploy agents but how to make them reliable enough to touch a customer's money.

"Saying that you're going to be efficient by putting a target in place is not the same thing as reimagining your processes through an agentic lens," Haus told ZDNet. He frames the shift as a redefinition of professional work — inside the IT department and across business lines.

"We are seeing that our traditional engineer, who was writing code in the past, can do more than just write code," Haus said. "And a product leader in the business, who traditionally was creating stories, can now write more code."

His playbook for making that real in a regulated environment comes down to three fundamentals: seamless services, secure products, and reliable outputs.

What platform does 200,000+ bank employees actually use?

The plumbing behind Haus's plan is LLM Suite, Chase's internal agentic platform, released in summer 2024. It exposes both frontier and open-source large language models to staff through a secure environment, and it handles the three jobs Haus expects every team to run: asking questions, reviewing documents, and drafting specifications.

Key features of the platform:

• Available to any employee at Chase, not only the engineering function
• Pulls in both closed and open-weight models behind a single interface
• Built to remove friction from the end-to-end process, not just one task

"It's a platform that anyone in the organization can use," Haus said. "We're looking at the end-to-end process and using the technology to remove the things that got in our way and to unleash our teams across the organization."

The customer-facing payoff is meant to be invisible. Chase routes inbound calls using machine-learning intent detection to reach the right person faster and runs similar systems mid-call to suggest next actions. A customer who calls about a card dispute should not need to know that an agent picked up the cue.

"The real magic is making things easier for a customer, so they don't even realize that the experience has been made better," Haus said. "It just works for them."

How does Chase keep agents secure in a regulated bank?

For Haus, security and privacy are non-negotiable inputs to the agentic rollout, not features bolted on at the end. As a bank operating under federal and state oversight, Chase holds emerging technology to the same bar it has applied to any model in production.

"We are thoughtful, and we follow the practices we've had in the past, where we can move quickly but responsibly," he said.

The same data-handling rules that govern traditional machine learning govern large language models. Consent and disclosure are explicit, not assumed.

"If you use data for something, we ensure our customers know," Haus said. "We would never use the data in an insecure way or against any of our privacy, compliance, or regulatory guidelines. Whether it's traditional machine-learning models or agentic technologies, those practices are critical."

The discipline cuts both ways. AI-enabled development tools help engineers flag bugs, surface adversarial behavior, and shorten the time between a vulnerability report and a patch. Haus pushes that posture on every team, not just on security specialists.

"The thing that I believe is incredibly important is continuing to make sure, outside of what we do to make AI work well, that we're focused always on our perimeter, ensuring that we are keeping our software updated, and that we are working on addressing vulnerabilities in a timely fashion," he said. "This proactive approach ensures that, as the world around us shifts, we can protect our customers."

Why are probabilistic outputs a banking problem?

Haus's third principle — reliable outputs — attacks the deepest weakness of agentic systems: they are statistical, not deterministic. The same model can return two different answers to two identical prompts, and the failure mode is rarely a clean error message.

The stakes differ by domain. A wrong answer in a recipe is a different brand of mayonnaise. A wrong answer in a payments workflow is a regulatory incident.

"You won't always get the same answer," Haus said. "This outcome matters less when you're looking for a recipe. If you end up with a different type of mayonnaise, it might be an issue. But when you want to do something with your finances, the output needs to be correct."

Chase responds with two moves. First, it keeps a human in the loop on most deployments — a pattern Haus called out explicitly. Second, it leans on engineers who can judge what an agent produced, not just whether it produced something. Evaluation, architecture, and the definition of "done" become the hard parts of the job.

"Thinking through and providing clarity on what must be true from an outcome when you are building software — from scale to the components that must be used to how it should be architected — are the hard parts," Haus said.

So what changes for the engineering workforce?

Haus is blunt about the implication. Coding is no longer the moat. Judgment is.

"I don't really hire engineers to write code," he said. "Now, I know that sounds weird because that's what I should hire engineers to do. But today, I hire engineers to know what code to write. There's a big difference. Until six months ago, you had to write the code. But now you don't have to because an agent can help you. This capability means a lot of the tedium that got in the way of what engineers really like to do goes away."

The reversal reframes the hiring bar, the career ladder, and the day-to-day work of every technologist at the bank. Patterns matter more than keystrokes. Architects matter more than typists. Domain experts who can specify the outcome — and recognize whether the model hit it — become the scarce resource.

"What are the patterns? How do we make sure those patterns can be followed so that, when our people build software, they're building the outcome the way they want, not just getting something the model is spinning out that isn't thought through holistically?" Haus asked. "That's the piece that I think is a fundamental that we need to capture. And you can do that in a variety of different ways in an organization, more than just using your talented engineers to write code."

What does this mean for every other bank?

Chase is among the first major U.S. banks to put a public stamp on the post-code engineer, and the timing matters. Rival retail banks are racing to stand up their own internal LLM platforms, often modeled on what they can observe from the outside. The questions Haus treats as settled — where agents sit in the operating model, who owns output quality, how shadow AI is policed — are still open at most institutions.

Two near-term pressures will force those choices. First, the talent market: engineers whose value is defined by typing speed will compete with agents that type faster for free. Second, the regulator's gaze: any bank that deploys a non-deterministic system against customer money inherits a duty to prove it works, and to prove it again every time the model changes.

Chase's bet is that the answer to both pressures is the same: redesign the work, train the staff, and treat the human as the verifier, not the writer. If the rest of the sector follows, the entry-level bank engineer of 2027 will look very little like the entry-level bank engineer of 2023 — and the agents that work beside them will be the least interesting part of the story.