What are the lessons from rolling out an enterprise AI assistant in fashion?
General-purpose AI assistants are the easiest AI to buy and the hardest to measure. Published rollouts and survey data point to a handful of repeatable lessons on scope, data access, training and proof.
KEY TAKEAWAYS Summary by the editors
- The main lesson from published enterprise assistant rollouts is that licences are the easy part: the value depends on defined use cases, access to the right data, training and a measurement plan.
- In March 2026, Marks & Spencer announced 11,000 Microsoft 365 Copilot licences for store managers and support centre staff, with reporting describing intended uses but no measured outcomes.
- McKinsey's 2026 survey found 80% of respondents say AI improved their individual productivity, yet only 37% attribute any EBIT impact to AI, so personal productivity does not automatically become company results.
- Under Article 4 of the EU AI Act, providers and deployers must take measures to support AI literacy of their staff, an obligation that has applied since 2 February 2025.
- Roles with a narrow, repeatable task and a clear data source are better rollout starting points than a company-wide licence drop.
The consistent lesson from enterprise AI assistant rollouts is that the licence is the smallest part of the project: results depend on choosing specific jobs for the assistant, giving it access to clean and permitted data, training people for those jobs, and measuring outcomes against a baseline. Fashion companies, with their fragmented product, sales and supply chain systems, face these issues in sharper form than most. This article draws on published rollouts and survey evidence, and is cautious where outcomes have not been reported.
What does a large fashion retail rollout look like?
One of the clearest public examples comes from UK retail. In March 2026, Marks & Spencer announced 11,000 Microsoft 365 Copilot licences for all store managers and support centre staff, according to reporting by Resultsense. The stated uses for stores include pulling together store-specific data such as sales performance and rota changes that previously required checking several systems, along with shift handover notes. At the support centre, the intended uses are meeting updates, trading summaries and recommendations from complex trading reports. M&S said it would run a training and development programme.
The same report notes that no measured outcomes had been published, only an anecdotal comment from one store manager, and that the proof would be whether the tools deliver measurable improvements in store performance and staff productivity. That caution applies to every company: a rollout announcement is evidence of intent, not of benefit.
Why do personal productivity gains not always show up in the accounts?
McKinsey's 2026 survey of 1,719 respondents found that 80% say AI has improved their individual productivity and 50% say it helps them make better decisions. Yet 37% attribute at least some EBIT impact to AI, essentially unchanged on the previous year. The survey also found that nearly three-quarters of high performers had fundamentally redesigned workflows around AI, against about one-quarter of other respondents.
The implication is that giving people an assistant changes how individuals work, but business results need changes in the process around them: who reviews, who decides, what gets dropped. In fashion, a merchandiser who drafts a trading summary in ten minutes instead of an hour has gained time, but the benefit is real only if the weekly trading meeting, the report format or the staffing assumption also changes.
What are the main lessons?
| Lesson | What goes wrong without it | Fashion example |
|---|---|---|
| Start with named use cases | Licences sit unused or are used for trivial tasks | Weekly trading summaries, supplier email drafting, line-sheet text checks |
| Fix data access first | The assistant cannot see sales, stock or product data and gives generic answers | Product attributes held in several systems with different codes |
| Train by role | One generic course; people do not apply it | Separate sessions for planners, e-commerce, wholesale sales and design |
| Set rules for what must not be pasted in | Confidential line plans or supplier prices enter the wrong tools | Pre-collection range data, costings, customer lists |
| Measure against a baseline | Anecdotes substitute for evidence | Hours spent on reporting before and after |
| Keep a human reviewer | Plausible errors reach customers or partners | Wrong size or composition text in a product page draft |
How should access to data and permissions be handled?
An assistant that can read internal documents inherits the quality of your permissions. If a shared drive is open to everyone by accident, the assistant can surface its contents to everyone. Review access rights on the sources you connect before switching the assistant on, and start with a narrow set of sources. Retail guidance from Anthropic, a model vendor, makes a similar point about fragmented systems across e-commerce, point of sale, inventory and customer data not being built to work together, and recommends building the foundation with stakeholder alignment and governance before scaling.
What does the EU AI Act require of staff training?
Article 4 of the EU AI Act requires providers and deployers to take measures to support the AI literacy of their staff and others using AI systems on their behalf, taking account of their knowledge, experience and context. According to the European Commission, the obligation has applied since 2 February 2025, no specific level or certificate is mandated, and enforcement sits with national authorities. Measures must be proportionate to the role. A fashion company rolling out an assistant therefore has a legal reason, as well as a practical one, to document who was trained, in what, and when.
Should you roll out to everyone at once?
Anthropic's retail guidance recommends selective, lower-risk pilots followed by scaling what works. The M&S example, by contrast, started with a defined population: store managers and support centre staff. The common thread is a bounded group with defined tasks, so that effects can be seen. A staged approach also lets you collect the first real questions people ask and adjust training before the wider group arrives.
- Pick two or three roles with repeatable, text-heavy or data-summarising tasks.
- Write down the baseline: time spent, error rates, turnaround.
- Review data access and permissions on connected sources.
- Train by role with real examples from your own work.
- Review usage and outcomes at six weeks and three months, then decide whether to widen.
- Publish short internal rules on what data may and may not be used.
The sensible expectation for an enterprise assistant is steady, modest gains in specific tasks, plus a clear view of where it does not help. Companies that set that expectation early are less likely to abandon the tool when early enthusiasm fades. They are also better placed to explain to finance, works councils and employees what the tool is for, which tasks it should not be used for, and how its results will be judged.
Frequently asked questions
What is an enterprise AI assistant?
It is a general-purpose generative AI tool, such as Microsoft 365 Copilot, deployed across a company so that staff can draft, summarise and query internal information. It is usually connected to company documents and systems under existing permissions.
Which roles in a fashion company benefit first?
Roles with repeatable, text-heavy or report-summarising tasks, such as trading analysis, supplier and customer correspondence, and product copy checking. Value depends on data access and a defined workflow, not on the role title alone.
Does the EU AI Act require training for staff who use AI?
Article 4 requires providers and deployers to take measures to support the AI literacy of staff using AI systems. No specific certificate or level is mandated, and the measures should fit the role and context.
How do you prove an assistant is worth the licence cost?
Measure a baseline before the rollout, such as hours on a specific report, and compare it after three to six months for the same task, allowing for seasonality. Usage statistics alone do not show business value.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
