How to start with AI in a fashion company: a 90-day plan
A practical three-month plan for fashion companies: choosing the first use cases, checking the data, running a pilot that proves something and deciding what to scale.
KEY TAKEAWAYS Summary by the editors
- A fashion company can get from first discussion to a tested AI pilot in about 90 days if it limits itself to two or three use cases with a clear business owner and a measurable outcome.
- MIT's NANDA initiative reported in 2025 that about 95 per cent of generative AI pilots it studied delivered little to no measurable impact on profit and loss, mainly because tools were not integrated into real workflows.
- Gartner predicts that through 2026 organisations will abandon 60 per cent of AI projects that are not supported by AI-ready data, so a data check belongs in the first month.
- In McKinsey's 2026 State of AI survey, nearly three quarters of AI high performers reported fundamentally redesigning workflows, which suggests pilots should change processes, not just add a tool.
- Basic governance from day one, such as a register of AI tools, rules for customer data and staff training, avoids costly rework later and supports EU AI Act duties.
The best way to start with AI in a fashion company is to pick two or three concrete business problems, check that the data for them exists, and run a short, measured pilot against a control group. Ninety days is enough to learn whether a use case works in your organisation, provided the scope stays narrow and someone in the business owns the result.
The plan below is written for brands and retailers without a large data science team. It assumes a sponsor in senior management, a small working group and a modest budget, and it deliberately avoids buying a large platform before the company knows what it needs.
Why do so many AI starts go nowhere?
Many companies begin with enthusiasm and end with a collection of disconnected experiments. MIT's NANDA initiative, in its 2025 report The GenAI Divide, found that about 95 per cent of the generative AI pilots it studied delivered little to no measurable impact on profit and loss, as reported by Fortune. The researchers attributed this less to the quality of the models than to a learning gap: generic tools were not adapted to specific workflows and did not improve with use.
Data is the other common failure point. Gartner predicts that through 2026, organisations will abandon 60 per cent of AI projects that are not supported by AI-ready data, and its 2024 survey found that 63 per cent of organisations either lacked or were unsure whether they had the right data management practices for AI. A 90-day plan that ignores data is likely to fail in the second month.
What should happen in days 1 to 30?
The first month is about choosing well. Collect candidate use cases from the business, not from vendors, and score them on value, feasibility and data readiness. Good first candidates in fashion share three traits: they repeat often, they have a clear measure of success and the data already sits in one or two systems.
| Use case | Success measure | Data needed | Typical readiness |
|---|---|---|---|
| Generating and translating product descriptions | Time to publish, error rate, conversion | Product attributes, composition, images | High if attributes are structured |
| Internal assistant for product and policy questions | Time saved per query, adoption | Manuals, policies, product data | Medium, depends on document quality |
| Size curve or replenishment suggestions | Stock-outs, excess stock | Sales by size and location, stock levels | Medium, often limited by lost sales data |
| Customer service reply drafting | Handling time, customer satisfaction | Past tickets, order and returns data | Medium to high |
| Order entry from emails and PDFs | Manual minutes per order, error rate | Example orders, product master data | High for repeat formats |
By day 30 the company should have chosen its use cases, named a business owner for each, written down the baseline (for example, how many minutes it currently takes to publish a product) and checked the data in a short, hands-on audit rather than a long report.
How should days 31 to 60 be used for a pilot?
The second month is the pilot itself. Keep it small and comparable: one category, one market, one team or a selection of stores, with a similar group working the old way as the control. Agree in advance what result would justify scaling and what result would mean stopping.
- Prepare the data for the chosen scope only, and record every correction you make, because that list becomes your data improvement backlog.
- Configure or buy the tool, and connect it to the systems people actually work in rather than asking them to copy and paste.
- Train the users in a short session that covers how the tool works, where it fails and when to escalate.
- Measure weekly against the baseline and the control group, including the time people spend checking AI output.
- Collect qualitative feedback from users, because adoption problems usually show up there first.
Human review should be built into the pilot. A product description that invents a fibre composition is a compliance problem, and a replenishment suggestion that ignores a store refit is a cost. Review rates can be reduced later once error levels are known.
How do you decide in days 61 to 90 what to scale?
The third month combines the final measurement with a decision. Compare the pilot with the baseline and the control group, calculate the full cost including licences, integration, review time and training, and decide whether to scale, adjust or stop. Stopping a pilot that does not work is a good outcome: it saves money and teaches the organisation something.
Scaling usually means changing the workflow, not just rolling the tool out more widely. In McKinsey's 2026 State of AI survey, nearly three quarters of high performers, the 6 per cent of respondents attributing at least 5 per cent of EBIT to AI, reported fundamentally redesigning workflows because of AI. For a fashion company that might mean that copywriters become editors of generated text, or that planners review exceptions instead of building every forecast by hand.
What governance do you need from day one?
Governance does not have to be heavy at the start, but it has to exist. A simple register of which AI tools are in use, by whom and with what data prevents surprises. Rules for customer and employee data, for disclosing generated imagery and for reviewing output before publication cover most early risks.
In the EU, the AI Act's provision on AI literacy has applied since 2 February 2025. Following the AI Omnibus published in July 2026, it requires providers and deployers to support the development of AI literacy rather than guarantee a specific level, according to the Future of Privacy Forum. Short, role-specific training for pilot users is a practical way to meet that expectation. Frameworks such as the voluntary NIST AI Risk Management Framework, organised around the functions Govern, Map, Measure and Manage, can provide a structure as the programme grows.
What mistakes should a fashion company avoid when starting with AI?
- Starting with a technology or vendor demonstration instead of a business problem.
- Running pilots without a baseline or control group, which makes the results impossible to judge.
- Spreading the effort across ten use cases instead of finishing two.
- Ignoring the time people spend checking AI output when calculating savings.
- Leaving IT, legal and works councils out until the end, when their concerns can stop a rollout.
- Signing long platform contracts before the company knows which use cases work.
After 90 days the company should have one or two proven use cases, a realistic view of its data gaps and a small team that knows how to run the next pilot. That is a more valuable starting point than an ambitious strategy document with nothing tested behind it.
Frequently asked questions
How long does it take to implement AI in a fashion company?
A focused pilot can be planned, run and evaluated in about 90 days. Scaling a successful use case across categories, markets or teams usually takes longer, because it involves integration, data clean-up and changes to workflows. Large programmes are built from a series of such pilots.
What is the best first AI project for a fashion brand?
Good first projects repeat often, have a clear measure of success and use data that already exists in one or two systems. Product description generation, order entry from emails and customer service reply drafting are common examples. The right choice depends on where your company loses the most time or money.
Do we need a data scientist to start with AI?
Not necessarily for the first pilots. Many early use cases rely on purchased tools or configured language models, and the critical skills are process knowledge, data stewardship and clear measurement. Specialist skills become more important when you build models on your own data, such as forecasts.
How much does an AI pilot cost?
There is no reliable industry figure, because costs vary with scope and tools. A business case should include licences or usage fees, integration work, data preparation, the time people spend reviewing output and training. Counting only licence costs usually understates the real investment.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
