Why do AI pilots stall in fashion, and how do you scale them to production?
Most generative AI pilots never reach scaled use. Why fashion pilots stall, from data and integration to unclear value, and a step-by-step route from proof of concept to production.
KEY TAKEAWAYS Summary by the editors
- MIT's 2025 study The GenAI Divide reported that about 95% of corporate generative AI pilots stalled before scaled adoption, attributing this mainly to a learning and integration gap rather than model quality.
- Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value.
- Gartner also predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 for similar reasons.
- Fashion pilots most often stall because they are run on clean sample data, outside real systems, without an owner, a baseline or a budget for the production phase.
- A pilot should be designed for production from day one: a named business owner, a measured baseline, live data, integration into the actual workflow and an agreed decision point.
AI pilots in fashion stall mainly because they are designed as experiments rather than as the first phase of a product: they run on tidy sample data, sit outside the systems people actually use, have no business owner and no agreed measure of success. Scaling requires the opposite: live data, workflow integration, a named owner, a baseline and a funded production plan agreed before the pilot starts.
How many AI pilots actually reach production?
The evidence suggests a minority. MIT's report The GenAI Divide: State of AI in Business 2025, based on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public deployments, found that about 95 percent of corporate generative AI pilots stalled at an early stage and never reached scaled adoption, as reported by Computing in August 2025. The lead author attributed this less to model quality than to a 'learning gap' between the tools and enterprise workflows.
Analyst forecasts point the same way. In July 2024 Gartner predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, because of poor data quality, inadequate risk controls, escalating costs or unclear business value. In June 2025 it predicted that over 40 percent of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.
Why do fashion AI pilots stall?
The reasons in fashion are usually practical and predictable.
| Stall point | What it looks like | Fix |
|---|---|---|
| Sample data | Pilot uses one clean season or one brand; production data has gaps, duplicates and inconsistent attributes | Test on a representative slice of live data early |
| No integration | Outputs sit in a separate tool and must be copied into PIM, ERP or e-commerce | Plan the integration path in the pilot scope |
| No owner | Innovation or IT runs the pilot; the business team that would use it is not accountable | Name a business owner with targets |
| No baseline | Nobody measured the current process, so benefits cannot be shown | Measure time, cost or error rate before starting |
| Cost surprise | Usage costs, licences or integration work exceed the pilot budget at scale | Model production costs before approving the pilot |
| Risk late | Legal, security or brand review happens after the demo | Involve governance at kick-off |
Seasonality adds a fashion-specific twist. A forecasting or assortment pilot may need a full season to show results, and if it starts at the wrong moment in the calendar it can miss the next buying cycle entirely, losing momentum and sponsorship.
How should an AI pilot be designed so it can scale?
A pilot that is meant to scale looks different from a demo. Before approving one, agree the following in writing:
- The business problem and owner: one sentence describing the problem, and the person whose targets improve if it works.
- The baseline: today's cycle time, cost, error rate or sell-through for the process in question.
- Success thresholds: the level of improvement that justifies production, and the level that means stop.
- Data scope: which real data sources will be used, and who fixes quality issues found.
- Integration path: where outputs will land in production (PIM, ERP, planning tool, service desk) and what that integration will cost.
- Risk review: data protection, IP, brand and EU AI Act considerations, signed off at the start.
- Decision date: a fixed point to scale, adjust or stop.
What changes between pilot and production?
Production adds work that pilots often ignore: monitoring output quality over time, handling exceptions, user support, access controls, cost tracking and updating models or prompts when products, policies or data change. For generative AI, someone must also maintain the reference content the system relies on, such as brand guidelines, care instructions or returns policies. Budget for this run cost from the start; it is a permanent operating expense, not a one-off project cost.
McKinsey's State of AI 2026 survey found that 44 percent of respondents said AI was scaling across their enterprise, up from 38 percent a year earlier, while the share attributing at least some EBIT impact to AI stayed essentially flat at 37 percent. The firms reporting the strongest impact were far more likely to have fundamentally redesigned workflows. Scaling a tool without changing how work is organised tends to deliver usage without value.
Communication also decides whether a pilot survives. The people whose work will change should see interim results, including the errors, and be able to shape the tool before it is rolled out. Pilots that are presented to users only once they are finished tend to meet resistance that no amount of model accuracy can overcome.
Should fashion companies build or buy to get past the pilot stage?
The MIT study reported that tools bought from external vendors or partners succeeded about two-thirds of the time, compared with about one-third for internally built systems. That does not mean building is always wrong; it means that internal builds carry more integration and maintenance burden than teams often expect. For common tasks such as product content or service replies, mature external tools may reach production faster. For proprietary decisions that rely on unique data, such as allocation logic, building may still pay off, provided the company has the engineering capacity to run it.
How many pilots should a fashion company run at once?
Fewer than most do. A small portfolio of three to five pilots, each with an owner, a baseline and a decision date, is easier to steer than dozens of loosely supervised experiments. Gartner's and MIT's findings both point to the same lesson: value comes from integrating a few use cases deeply into real work, not from running many disconnected trials.
Frequently asked questions
What percentage of AI pilots fail?
MIT's 2025 report The GenAI Divide found that about 95 percent of corporate generative AI pilots stalled before scaled adoption. Gartner predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025. Definitions vary, but both point to most pilots not delivering scaled value.
Why do AI proofs of concept fail?
Common causes are poor data quality, lack of integration into existing systems, unclear business value, escalating costs and risk controls added too late. In fashion, seasonal timing and inconsistent product data are frequent additional obstacles.
How long should an AI pilot run?
Long enough to measure the outcome against a baseline, which for content or service use cases may be weeks and for forecasting or assortment may be a full season. The end date and decision criteria should be fixed before the pilot starts.
What is the difference between an AI pilot and production?
A pilot tests whether a use case can work on limited scope. Production means the tool runs on live data inside the real workflow, with monitoring, support, access controls, cost tracking and a business owner accountable for results.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- Computing: MIT report: 95% of corporate generative AI pilots fail to deliver returns
- Gartner: Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept by End of 2025
- Gartner: Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
- McKinsey: The state of AI
