AI won’t fix a messy ERP: ASOS’s £130m stock write-off proves it

AI gets a lot of attention for what it might do - sharper predictions, smoother shopping journeys, quicker answers. But it can't rescue a shaky foundation. ASOS offered a quiet reminder of that. Sales were sliding, pressure was mounting, and out came a polished AI "stylist" meant to tempt shoppers back. The idea wasn't the problem. The groundwork beneath it was.

Behind the scenes, the real issue was familiar to most retailers: an ERP straining under messy product data, muddled returns codes, and stock updates that arrived a beat too late. When those foundations slip, AI doesn't soften the blow. It speeds it up. The system starts to recommend items that aren't actually in stock, forecast demand from unreliable numbers, and make delivery promises the warehouse can't keep.

And as every retailer knows, peak season puts those cracks under a spotlight. For smaller and mid-sized brands, the real advantage isn't the clever model - it's the everyday discipline beneath it: cleaning the item master, tightening returns logic, getting inventory signals consistent and on time. Once that foundation is steady, AI finally has something solid to work with. It stops amplifying the chaos and starts amplifying the craft.

AI algorithms cannot solve bad ERP data

Retailers have rushed to bolt AI onto the front of eCommerce: stylists, chat assistants, "complete the look" modules and dynamic sort orders. The promise is real. But the minute those tools pull from inconsistent fields, duplicated SKUs or lagging inventory feeds, they start to make the site look unreliable. A recommendation carousel that surfaces a size that sold out yesterday isn't helpful; it's a customer-service problem waiting to happen.

The ASOS write-down is a cautionary tale in plain numbers. If a FTSE-listed pure-play fashion giant can be forced into a nine-figure stock clean-up, smaller players should take note. AI did not create the overhang, but without tight ERP discipline, any machine-learning layer draws the wrong conclusions faster and shows them to more shoppers.

Where ERP data hygiene breaks - and why AI magnifies it

Three weak spots show up again and again in UK retail operations:

1. Item master sprawl. Product titles, attributes and category tags diverge across seasons and suppliers. One team calls it "mid-wash," another "light blue," a third abbreviates it to "Lt Blu." AI models trained on behaviour signals will treat these as different items unless the ERP standardises them. The result: fragmented analytics and thin, misleading training data.
2. Returns codes that don't explain reality. "Damaged," "didn't fit" and "changed mind" aren't granular enough to diagnose faults or size curves. If fit feedback is muddled, personalisation engines keep pushing silhouettes with a high return propensity, artificially inflating "engagement" while harming margin.
3. Inventory signals out of sync. Even a 10–15 minute delay between warehouse adjustments and website availability can undermine trust at peak. AI-driven back-in-stock alerts or "selling fast" banners mean little if the source feed lags - or if safety stock buffers aren't enforced centrally.

AI thrives on patterns. But if your foundational data is inconsistent, you're asking the model to learn from noise. It will still learn - only the lessons will be wrong.

"ERP first, AI second" is not anti-innovation - it's pro-outcome

There's a narrative that getting your ERP in order is a slow, back-office tax that delays customer-facing innovation. In practice, the opposite is true. Teams that invest in ERP hygiene see conversion lift precisely because front-end experiences stop breaking. Product findability improves. On-site search returns fewer dead ends. Recommendations look smarter not because the algorithm changed, but because the inputs did.

It's also cheaper. Every retailer is being asked to do more with less. Before you sign a new AI personalisation contract, ask whether you've wrung out the free gains sitting in your eCommerce ERP. Standardising attribute taxonomies, enforcing canonical naming, and aligning pack-size and colour codes cost far less than another front-end experiment - and they make every experiment work better.

A practical clean-up plan for the peak period (and beyond)

Editors hear "data hygiene" and fear a multi-year IT saga. It doesn't have to be. Four pragmatic moves will stabilise the foundation in weeks, not quarters:

1) Lock the language.
Create a controlled vocabulary for the item master: style, colour, fit, fabric, and care fields with allowed values only. Ban free text in those fields. Build validation rules that reject new SKUs unless they conform. If you sell via marketplaces, map the vocabulary to each channel's taxonomy and bake that mapping into your PIM or ERP extension.

2) Reset the returns matrix.
Rework reason codes to be precise and mutually exclusive (e.g., "too small - waist," "too small - length," "transparent fabric," "quality - seam failure"). Tie each code to a remediation action: size-curve rebalance, imagery update, product-page copy change, or supplier escalation. Send those signals back into recommendation suppression lists so the site stops pushing brittle items.

3) Shorten the inventory hop.
Audit the path from warehouse management system to ERP to eCommerce platform. Remove transforms and batch jobs you don't need. Move to event-driven updates or sub-five-minute polling. If that's not possible, expose a "last updated" timestamp to merchandising and customer-care teams so they can judge reliability at a glance.

4) Make data stewardship a habit, not a project.
Assign named owners for each high-risk field (size, colour, availability, lead time). Measure error rates weekly. Publish a simple scorecard - % SKUs conforming to taxonomy, inventory latency, returns-code usage mix - and put it on the ops stand-up agenda. What gets measured gets fixed.

What good looks like on the storefront

When ERP hygiene improves, the change is visible to shoppers in everyday moments:

• Search actually understands intent. "Blue high-waisted wide-leg jeans" returns the right set because the taxonomy is consistent, not because the search engine got smarter overnight.
• Recommendations stop hallucinating stock. The "complete the look" block only shows sizes you can ship today, and quietly removes styles with high return propensity until issues are addressed.
• Delivery promises match reality. Lead times reflect warehouse cut-offs and carrier capacity, so checkout no longer needs to plaster on disclaimers when storms hit.

Internally, trading teams gain confidence. They trust the numbers in the morning dashboard because the definitions behind them are stable. That trust encourages intelligent experimentation. Suddenly, AI pilots - size guidance, returns prediction, dynamic bundles - have clean ground truth to learn from.

Common objections - and how to answer them

"We can't pause innovation while we refactor the ERP."
You don't need to. Prioritise the dozen attributes that most influence findability, fit and fulfilment, fix those first, then iterate. This is a rolling program, not a freeze.

"Suppliers won't change their data feeds."
Fine - normalize at the edge. Build an ingestion layer that maps supplier terms to your vocabulary and rejects anything non-compliant. The pain is upfront; the payoff is lasting.

"Our AI vendor says their model handles messy data."
To a point. Most models can smooth inconsistencies; none can conjure stock that isn't there or decode a vague returns code into actionable insight. Garbage-in still means garbage-out, just faster.

Why this matters now

UK consumers are value-sensitive and impatient. A single poor experience - an out-of-stock at checkout, a mis-sized item pushed by a "smart" widget - nudges them to a competitor. As promotional calendars compress into fewer, louder peaks, the cost of small data slips multiplies. That's why the lesson from the ASOS write-down resonates beyond one brand: at scale, small misalignments turn expensive quickly.

The upside is equally clear. Retailers who put ERP hygiene first get to redeploy AI where it shines: understanding cohorts, optimising markdowns, predicting returns likelihood, and tuning content for conversion. None of that requires a gimmick. It requires trustworthy inputs.

The bottom line

AI is not a plaster for structural problems. If your ERP is messy, AI will amplify the mess - faster recommendations of the wrong items, slicker experiences that promise what your warehouse can't deliver. The fix isn't glamorous: standardise the item master, sharpen returns logic, and speed up inventory signals. Do that, and the clever stuff starts to work as advertised.

ASOS's stock write-off is the headline-grabber, but the core message is operational and universal: ERP first, AI second. Clean the foundation, then let the algorithms do their best work.