Retail versus finance: How genAI coding strategies diverge

An code analysis vendor on Tuesday published a study of the generative AI (genAI) coding differences between its retail and finance customers, revealing that the retail approach is much more aggressive, whereas finance has been working on the technology longer.

“Retail is pushing genAI into production faster. Finance is still experimenting. Retail companies are embedding genAI at 2.1 times the rate of financial services, based on the proportion of repositories that include genAI components,” said the study from Apiiro, whose clients include Colgate-Palmolive, Shell Global, BlackRock, and Rakuten Group.

The analysis, based on an examination of more than 100,000 code repositories using Apiiro’s Deep Code Analysis tool, showed how differently the two verticals are approaching genAI coding strategies. It noted, “61% of retail genAI repositories show active development, based on commit activity and contributor engagement. In financial services, that number [is] 22%. Retail teams are moving genAI projects through the build-test-ship cycle, while many finance teams remain in slower, more siloed experimentation phases.”

Much of the difference, the report said, comes from the tasks each vertical is pushing. “Retail teams are using genAI to power real-time, customer-facing features like recommendation engines and automated support. With shorter feedback loops and direct revenue impact, the incentive to ship is constant. Financial institutions, by contrast, operate under heavier regulatory scrutiny. Their genAI work is more cautious, often confined to internal systems and it shows in the development patterns.”

Financial services organizations, it said, have been working on genAI projects much longer, which is reflected in the ages of their repositories; the average age of a finance organization’s genAI repository is 688 days, significantly older than retail’s 453-day average.

Jason Andersen, VP and principal analyst at Moor Insights & Strategy, mostly agreed with the report’s findings and said the details matched his observations with clients in those two verticals.

He was intrigued by the age difference between the two sectors’ genAI repositories, but only because he would have guessed that retail would have been more of a laggard.

With finance’s “688 days, that’s roughly two years, which is just when most of the early (genAI) models started coming out,” Andersen said. “That makes sense because finance historically trends faster. They [understand] data better. I am amazed that it was 453 days for retail. I thought it would be much less. I think retail is moving even slower than that.”

The report said that retail genAI use cases typically involve customer-facing personalized product recommendations, automated support, and tailored promotions.

“These systems rely on real-time, user-specific context — and that means direct access to sensitive data,” it said. “In finance, genAI usage remains more siloed, less client-facing: pilot programs, internal assistants or data-abstracted training scenarios. Regulatory pressure plays a role, but so does engineering culture: [finance] pipelines are less often wired directly into live user data.”

Moor’s Andersen said that the industry priorities of the financial vertical means that financial IT “can be a lot more experimental” than its retail counterparts.

Financial “has more means, more money. That entire [financial] industry is based on beta so they are well primed for [genAI experimentation].” 

Retail IT attacks these developments the same way that it always has. “They look at every automation move the same since the tractor,” Andersen said. “Retail IT asks, ‘How am I going to use this to increase margins?’ And financial is looking at [genAI] for innovation. They are asking, ‘How do I create new products?’”

Another area the report explored was tool usage. 

“Financial services teams use a wide range of genAI tools, including OpenAI Client, LangChain, and LiteLLM. Their projects span multiple model types and dataset formats, which is a sign of active experimentation across varied use cases,” the report said. “Retail, by contrast, has converged on a smaller, tighter stack: OpenAI Python SDKs and LiteLLM dominate. These tools feed into high-leverage, customer-facing use cases like product recommendations and personalized search.”

Retail’s use of fewer tools, the report said, has the benefit of sharply accelerating operationalization. 

“Fewer tools mean fewer integration points, tighter pipelines, and more repeatable patterns, [whereas] finance’s broader stack creates fragmented risk surfaces and steeper governance complexity. What it gains in flexibility, it loses in consistency.”

Maman Ibraham, principal partner at the EugeneZonda consulting firm, had a more succinct take on the genAI tool usage in financial. 

“Having 20 genAI tools doesn’t make you innovative,” Ibraham said. “It makes you ungovernable.”

To mitigate risk, Apiiro recommends that each vertical calibrate its genAI coding strategies differently, based on its environment. “In retail, start with data mapping, access control audits, and early-stage static analysis to catch issues before deployment. In finance, prioritize secrets detection, dependency hygiene, and reviewing whether dormant genAI projects should be refactored or retired,” it said.

More on gen AI coding:

Developers set the pace for genAI tools adoption

What we know now about generative AI for software development

How generative AI rollouts fail, and how to fix them