How Soon Could The Verticalisation Of AI Come For Fashion?

A modern computer is humankind’s most versatile tool. The same device, running the same operating environment, can be used to send a stupid GIF to a group chat, file a lawsuit or an IPO, research and write the Great American novel, watch films, edit films, compose music, make art, design clothes, or intelligently arrange their pattern pieces to optimise material yield in a numerically-controlled manufacturing operation.

No matter the form factor (desktop, laptop, table, phone or face-mounted), a computer and an OS is such an adaptable combination that those two things are basically the prerequisites for all knowledge work, and for then translating that knowledge work into physical action through the orchestration and direct instruction of programmable controllers.

But that general-purpose nature of computers has made for a very complex and malleable surface that has at once opened the door to automation and resisted it.

On the one hand, it’s now straightforward for a technology company to drop an AI agent onto your computer that, given the proper permissions, can interact with the entire GUI and filesystem and take on literally any task that a human being might perform with the same degree of access.

This was, in fact, the entire thrust of yesterday’s launch of GPT-6 Astra, OpenAI’s newest frontier model. The promotional video showed a range of people talking to that model and asking it to perform 2D-to-3D design workflows, website-building, spreadsheet-wrangling, second-hand-selling, food-ordering, contract-drafting and more, all inside macOS.

This is the new, second-order effect of the consolidation of our work and personal lives into computers. It’s unambiguously true that more music has been composed, more films have been made, and more valuable art (across every medium, fashion included) has been created because users have been able to do those things on the do-anything devices they already owned, instead of needing to buy instruments, editing panels, canvases or fabrics. But it’s also true that these activities are now within the reach of AI as a direct consequence of them all coexisting in the same ubiquitous environment.

Or, to simplify it: because we do everything on computers, other entities that we give access to our computers can theoretically do everything as well.

They might not necessarily do a good job of it, though, for the same reason. The computer has become the funnel through which every domain has been pushed, but not every computer user is a polymath who plays and works across them all. 

Just because someone here at The Interline can open Logic Pro side by side with After Effects doesn’t mean that that single person can make music just as well as they can make motion graphics. And these are both creative tasks with some crossover; the difference becomes more obvious if we consider that one computer user in a coffee shop might be conducting research into MRNA vaccines while another is designing flyers for their restaurant.

The homogeneity of the computing environment is what makes the kind of robotic process automation that the frontier of “computer use” represents viable. But the heterogeneity of the tasks that people can perform on a computer is what renders full-spectrum task automation, across every domain, difficult to accomplish.

This is why a lot of effort is going into verticalising software – and in particular AI. This week, Anthropic (which also released its latest frontier model a few days ago, with far less emphasis on computer use, but far more on another area we’ll talk about in a moment) announced that it was launching a set of “blueprints” for retailers that would enable them to help prepare for the upcoming holiday seasons.

It’s not immediately clear what these blueprints actually include, but The Interline’s estimation is that these represent more turnkey ways to build natural language agents for product recommendations, inventory look-ups, and marketing. All of which are capabilities it’s been possible for brands and retailers to build for themselves, using models from Anthropic (or other sources!) and agent frameworks, but which are more than likely logical targets for pre-assembling based on best practices and reusable components. 

This is part of a wider push from every frontier AI lab to offer pre-packaged setups to specific industries such as finance and legal. And while we previously predicted that the same industry-specific packaging would eventually be done for fashion, these blueprints seem like a stopgap solution rather than any seriously big swing towards optimising Claude for retail use cases.

Nevertheless, this hasn’t stopped companies from using fashion workflows in their marketing materials for general purpose computer use. If we look back at the aforementioned launch video for Astra, the presentation workflow shown involves a designer developing ways to showcase their rainwear collection to potential retail partners, with the LLM taking on the lion’s share of the work.

Notably this is not an example of AI actually designing rainwear, which would involve computer use of a very different and more verticalised kind (2D and 3D design tool-use). But if early indications of Astra’s ability to model in Blender are any indication, this may not be too far away.

For all this push and pull between general-purpose software and verticalised processes, though, there’s a headline from the end of last week that’s arguably more indicative of how AI could end up impacting the way fashion works.

On the 27th August, Anthropic unveiled its new Model Hardware Standard, which is a shared specification that allows AI agents to control physical devices. The MHS is being piloted and demonstrated in laboratory settings, and there’s currently no suggestion that cut-make-sew factories are on the radar, but Anthropic’s previous standard (the Model Context Protocol, or MCP) has taken less than two years to become the de facto layer through which language models have become capable of controlling software.

As a case in point, where Astra and other computer use models are using Blender, or other 3D packages, they’re doing so using MCP. But the same standard is also how Claude and other language models manipulate data from a massive range of other systems, from accounting to email.

So while there’s nothing to suggest right now that Claude is preparing to try and drive cutting machines, or the other specialised hardware that takes up space on the typical shopfloor, MCP does provide us with at least an indication that complex setups (like those seen in working laboratories) can fall quite quickly to a common standard.