The Fashion Jobs Where AI Shines, and Where It Struggles

This February, Gap Inc. launched an internal application called GapGPT, a single portal for employees to access different large language models.

It routes queries to the best tool depending on the need: Copywriting jobs are directed to OpenAI’s ChatGPT, image-generating requests go to Google’s Nano Banana and coding tasks are sent to Anthropic’s Claude Code.

But while Gap trusts LLMs enough for these jobs, it isn’t ready to rely on them for anything involving complex math.

“You need to make sure that you’re using LLMs for what LLMs are good for, and using classic algorithms for what they’re good for,” said Sven Gerjets, the company’s chief technology officer. “LLMs aren’t classically good at math, so you want a different method for that.”

Over the past few years, fashion businesses have raced to put LLMs to work, hoping they can help speed up a range of otherwise time-consuming jobs and boost worker productivity and efficiency. But getting the most out of LLMs isn’t as simple as throwing them at any situation that comes along. Even as they continue to advance, they still have notable weaknesses — beyond the issues with math, there’s also the tendency to make up information, known as hallucinations.

For workers, it’s vital to know where and how to apply the technology. In a 2023 study, researchers coined the phrase “jagged technology frontier” to describe the uneven capabilities of AI, where it shines in some tasks and can improve productivity, yet it stumbles in other areas of equal difficulty.

“Outside the frontier, AI output is inaccurate, is less useful, and can degrade human performance,” they wrote. “AI assistance improves human performance only for tasks within current AI capabilities — within the jagged technological frontier.”

Three years later, these challenges persist, according to Ethan Mollick, one of the researchers involved in the study. The takeaway, however, isn’t that fashion companies need to separate tasks into those AI should or shouldn’t touch. What’s necessary is to understand the capabilities and shortcomings of any type of AI, allowing you to leverage the former and circumvent the latter.

Fashion businesses that are effectively able to use AI can find it a differentiator in a competitive market, according to UBS analyst Jay Sole, who wrote in an April 6 note to clients that companies with stronger AI adoption are seeing greater workforce productivity.

“It’s a big deal,” Sole said in an interview. “It’s something that can help a business improve its operations.”

Inside and Outside the Jagged Frontier

Areas where LLMs are already helping fashion companies tend to make use of their aptitude for parsing and generating natural language, or producing high-quality imagery. Designers are using AI to quickly visualise ideas, while marketers are creating copy and images to use on brand websites and in emails. During a virtual roundtable for press last month, Graeme Smith, senior vice president of engineering and science at Zalando, said 90 percent of the online retailer’s on-site marketing content in December was produced with generative AI.

Companies are also using LLMs to streamline laborious duties, such as compiling and analysing reports. Kelly Miely, a partner at Deloitte focused on buying, merchandising and supply chain, said an area where they’ve shown quick progress is doing in-season analysis and providing guidance on actions a company should take.

“It’s a task that typically took a huge host of admins, a lot of time and effort on a Monday morning, to gather all of the performance data, look at the competition, look at how they’ve performed, look at the ranges in different places and then understand what insight is there,” she said. “That is an area where you’re seeing the LLMs really help, because they’re able to generate the report, they’re able to generate the insights around it.”

What companies aren’t doing, she added, is using the LLM itself to crunch that performance data and do scenario modelling. That work continues to be done by machine-learning algorithms that are more reliable at performing complex calculations.

There are other jobs that companies should be cautious of entrusting to LLMs, such as financial auditing, completing regulatory filings and anything that requires the output to be 100 percent accurate. Even in the tasks they are being used for, LLMs are prone to making mistakes.

There’s still debate, for example, over how useful they are for coding. Many developers have rapidly adopted them for the job, and they’re effective enough to have given rise to what’s called vibe coding, where a user can create software by telling an LLM what they want in natural language and have it generate the code. That code still often requires debugging, however.

Even so, LLMs appear to be getting better at writing code. A 2025 study, which originally found using AI slowed down experienced developers who had to spend time correcting its output, recently reversed its conclusion, saying it now likely speeds them up. The caveat, however, was that because so many software engineers now use AI, it may have biased the results.

Still, a report by The Guardian last month included testimony from several Amazon workers, including developers, who said they were being pushed to use AI even though it hurt their productivity. Amazon said it had hundreds of thousands of employees in different roles using AI to learn what works and what doesn’t for them, and that different employees had different experiences, with many finding value in using AI.

The Frontier’s Fuzzy Borders

Those varying experiences point to one of the challenges of integrating LLMs into corporate workflows: It can be hard to know where they’ll help and where they’ll be a hindrance.

In the study of AI’s jagged frontier, the researchers pointed out that a few different factors make it difficult to guess ahead of time what the advantages and downsides of using generative AI will be. The technology displays a number of unexpected capabilities that keep progressing, which can make it very useful for knowledge workers, but it can also produce incorrect results that users will find believable. “The advantages of LLMs, although substantial, are frequently unclear to users,” they wrote.

Companies are navigating these challenges in a variety of ways. One common method is having an actual human review any output. When using AI for copywriting, for instance, someone should still be reading and editing that text.

But there are other ways companies can keep LLMs in check.

“When we’re creating workflows, it’s an architecture of capabilities we’re bringing together,” said Gap Inc.’s Gerjets.

A tool could have an embedded LLM performing some jobs but will call out to a separate algorithm to perform any mathematical calculations. For any uses of LLMs involving interactions with customers, Gap will have an AI agent generate a response and use another agent to validate that output before it’s passed on to the customer.

Gerjets said keeping a human in the loop is important, but you also don’t want a lot of variation in the LLM’s output that the human would need to correct. AI can help minimise those issues, too.

“In coding, for example, we’ll have an agent code, we’ll have another agent validate the code and that agent will tell the coding agent to go do other things if it finds issues,” Gerjets said. “You create this reinforcement loop that really allows you to manage around some of the things that they’re good at and some of the things that they aren’t good at.”

Companies would be unwise to write off LLMs entirely because of their flaws, but they also need to have a clear understanding of their vulnerabilities and how to mitigate them.

That’s another area where humans remain indispensable: Their judgment is critical to know what LLMs can and can’t do.

Want to dive deeper into an insight from this article? Check out The Brain of Fashion, BoF’s new generative AI tool where you can unlock BoF’s technology archive with a single question.