Why most fashion AI projects never make it out of the pilot phase

Somewhere in a fashion business, there’s a planner staring at two numbers that should match but don’t. One is from the buying plan. One is from finance. Both are meant to represent the same thing (committed spend), but neither is obviously wrong or right.

This is usually the moment someone says, “Once we get AI in, this stuff will sort itself out.”

Yeah, probably not.

According to The State of Fashion 2026, up to 90% of AI initiatives never scale beyond experimentation. Pilots often impress in isolation but quietly stall when they touch the real operating model, and cracks start to show.

Where things start to go wrong

The issue is rarely down to ambition. Teams already operate at maximum capacity, juggling late suppliers, volatile costs, compressed calendars and consumers who change their minds faster than production cycles allow.

The crux of the issue is that AI is often dropped into environments that already work around structural cracks.

Think about a scenario where a landed cost changes. It might start with a tariff increase, a sudden freight spike, or a supplier quietly raising their price, but the impact rarely stops there.

The information arrives somewhere, maybe procurement, but not everywhere.

Merchandising then works from old margin assumptions, finance sees the impact later, and allocation makes decisions anyway, simply because they have to.

By the time the season closes, everyone understands what has happened, but it’s too late to do anything about it.

This is the first mile problem, whether it’s labelled that way or not. Decisions that shape outcomes are made before products reach the warehouses, and to make things worse, they’re made inside disconnected systems, each with its own version of the truth.

So, when AI is asked to forecast demand, make correct inventory decisions, or identify ways to preserve margin, it’s ultimately being fed a picture of the business that is incorrect.

The quiet reason why AI gets stuck

AI pilots typically work because they rely on clean datasets, defined scopes, and a controlled environment. Everyone involved agrees on what “good” looks like.

But most fashion organisations do not have squeaky-clean data, clear scopes, or stable operating environments. The conditions are simply imperfect.

Scaling an AI model means deploying it inside the messiness of day-to-day fashion ops. That means late deliveries, split shipments, stepped buying, mid-season assortment changes, wholesale commitments that can’t be broken and finance trying to close as best they can.

AI initiatives don’t fail because of people, but because organisations rarely have a single operational spine that feeds correct data into the models.

What tends to happen instead is that teams keep working around the system until their core ERP becomes a reporting layer rather than a place where decisions actually happen.

AI ends up analysing the same contradictions and messy data people already wrestle with.

Fashion’s pain is commercial, not conceptual

When foundations are weak, the consequences are almost always financial.

A season can look healthy on the surface while quietly bleeding underneath. Stock builds in locations where demand never arrives, top-performing items sell through too early, and cash gets trapped in inventory just as freight bills, duties and markdown pressure start to peak.

By the time leadership sees the full picture, the season is already closing. The usual response is to push stock harder, discount faster, protect relationships where possible, and accept that margin has taken the hit this time.

Again, none of this happens because teams don’t understand their jobs. It happens because the system they work with can’t keep decisions consistent and aligned as conditions shift.

AI, in that context, is nothing more than insight without execution and prediction without control. It might explain what went wrong, or even what’s likely to happen next, but it can’t intervene early enough to change the outcome.

Why agentic ERP only works after the hard part

This is where agentic ERP actually earns its keep and where it’s most misunderstood.

In case you’re unfamiliar, agentic ERP refers to ERPs that use embedded AI agents to monitor data, surface issues, and automate routine actions directly inside core business workflows.

If planning, buying, sourcing, allocation, and finance are connected, agents stop being promises for the future and can deliver tangible outcomes in the here and now.

They can flag margin erosion before commitments are locked, reconcile continuously instead of monthly, and adjust workflows when variables change, without forcing manual intervention.

But this only works when the system underneath already knows how the business fits together.

Microsoft D365 Business Central is one of the best examples in the market right now. Since Copilot and agents are embedded into the ERP itself, they’re also embedded into the operational backbone where decisions are made and executed.

Everything lives in one system rather than across disparate tools and spreadsheets, which means AI actually works with data that reflects reality.

The brands that scale AI do something unfashionable first

With all the hype around AI and all the tech vendors promising “cutting-edge AI efficiency,” many businesses have forgotten that stabilising the back office is priority number one.

This is the difference between brands that scale AI and those that stall in the pilot stage.

Smart brands accept that resilience isn’t built at the warehouse door or the last mile, but upstream in how plans are connected, costs are modelled, and changes propagate throughout their organisation in real time.

Once that’s in place, AI stops being something you “roll out” and starts being something that quietly compounds. It’s neither flashy nor sexy, but it is effective.