6 November 2026International edition
Vol. I · No.
6 November 2026
AI IN FASHION
DAILY
The daily briefing on AI in the fashion business
Where fashion meets artificial intelligence.
Strategy, Data & Regulation · Stories

How to run a 90-day AI pilot that proves value or stops cleanly

Pilot design for wholesale use cases: scope definition, baseline metrics, control groups, decision gates at 30/60/90 days, and exit without sunk cost fallacy.

a long blue bridge
Photo: Alberto Rodríguez / Unsplash

KEY TAKEAWAYS Summary by the editors

  1. Up to 90 percent of fashion AI projects are stuck at pilot stage, constrained by structural and cultural barriers.
  2. A 90-day pilot splits into three 30-day phases: scope and baseline, run in parallel with control group, and validate against agreed thresholds.
  3. Baseline metrics such as order accuracy, margin variance and time per task must be captured before the pilot starts, not after.
  4. Exit criteria written before the pilot begins prevent sunk cost fallacy and allow clean stop decisions at each gate.

Why do most fashion wholesale AI pilots fail to scale?

Up to 90 percent of transformative AI projects are stuck at pilot stage, constrained by structural and cultural barriers. An August 2024 survey from McKinsey found that 64 percent of retail leaders had conducted AI pilots, yet less than 20 percent of AI pilots cross the threshold into enterprise-scale production. The gap is not technical. A pilot is an experiment with a decision attached, and one without a written stopping rule cannot produce a decision, so it produces an extension instead. In wholesale, where order volumes, margin and speed define competitiveness, a pilot that drifts without measurable proof wastes both capital and credibility.

Fashion wholesale is regaining momentum, order lead times are getting shorter, brands are developing tighter collections featuring more evergreens, and both retailers and brands are increasing their use of AI tools. Widely used digital wholesale platforms like Joor and Lightspeed's NuOrder have launched AI tools to help brands strike the right inventory balance and optimise their buys. The challenge is not whether AI can improve pricing, forecasting or order management. The challenge is proving it did, in a time window short enough to course-correct or stop.

What belongs in a 90-day wholesale AI pilot scope?

The 90-day structure forces discipline. It splits naturally into three 30-day phases: scope, run, and prove. Each phase has exactly one job and one exit gate. 90 days is the standard for enterprise AI pilots: 30 days for scoping and data readiness, 30 days for build and integration, 30 days for evaluation and go/no-go decision. In wholesale, the scope must be narrow enough to control variables but significant enough to matter financially.

Suitable use cases include demand forecasting at style and colour level, pricing optimisation for wholesale accounts, order accuracy improvement, or reorder prediction. For each product category compute five numbers: revenue, gross margin percentage, margin variance across customers, transaction count, and average line value. Margin variance is the one that matters. Categories with wide unexplained variance are where pricing work pays first. Pick exactly one use case. Define the success metric before you start, numerically, with today's baseline: lift gross margin on category C by 80 basis points by June 30, not improve pricing.

Scope definition also includes what the pilot will not do. Complete security review, provision access for the pilot group only, and produce a one-page scope document naming what is in scope, what is out of scope, who the three roles are, and what the success metric is. This prevents scope creep and ensures the pilot remains reversible.

How do you capture baseline metrics before the pilot starts?

Baseline measurement is the step most teams skip, and the reason most pilots cannot prove their value later. You have a reference KPI with a current baseline measurement. Every AI build needs a number agreed before launch, measured against a baseline from the same season last year. These are the metrics that tell you whether a deployment is paying or just running. In wholesale, baselines include order accuracy rate, margin realised per category, forecast error at SKU level, and time per order processed.

Baseline measures include conversion, return rate, stated return reasons, customer questions and full value resale of returns. Record a baseline for the selected products, then introduce the tool to only part of the selected range or audience so that performance can be compared with a similar control group. Validation against baseline: compare the pilot to the manual process over the same periods. For pricing, measure realised margin on the treated categories against a control group of similar untreated categories, which is the only comparison that survives scrutiny.

Why are control groups essential in wholesale pilot design?

Control groups isolate the effect of the AI from seasonal variation, market shifts and other concurrent initiatives. To test the hypothesis that model recommended prices are better than baseline prices, two user groups were created: Set-A (control group) was shown the baseline prices, Set-B (treatment group) was shown model recommended prices. These sets were created using a random assignment from the set of live users. 50 percent of the total live users were assigned to Set A and the rest to Set B. Without a control, it is impossible to distinguish AI impact from natural fluctuations.

Measure return on investment (ROI) in each store, in the KPI families the AI should move, against a baseline. Pilot locations should beat both their own past and a matched control group. In wholesale, the control can be a subset of accounts, product categories, or time periods. Two successive live A/B tests were performed with over 300,000 customers per group. Using controlled and live A/B test settings, customers were randomly divided into two groups; a control group that received no size advice and a test group with personalised advice. The control must be truly comparable, or the comparison is worthless.

What happens at the 30, 60 and 90 day decision gates?

Each gate has a specific purpose. At day 30, confirm the pilot is instrumented well enough to be scored at all, since most pilots that cannot be evaluated at day 90 were never set up to capture the right data. At day 30, the question is not whether the AI works. The question is whether the pilot is producing the data needed to decide at day 90. This includes confirmation that baseline capture is complete, control and treatment groups are defined, and the metric can be calculated.

At day 60, the pilot should be running in parallel with the existing process. Ship small, then expand the sample. Week 5: the AI does the task alongside the human, output marked but not used. Week 6: the AI output becomes the first draft the human edits. Early results may be visible, but the gate tests readiness for full evaluation, not final performance. Issues such as integration gaps, user resistance, or data quality problems surface here. If the pilot cannot run reliably for 30 days, it will not scale.

At day 90, the decision is binary: scale, stop, or redesign with a new pilot. An agentic AI pilot should be scored at the 90-day mark on five weighted dimensions totaling 100 points: Outcome Integrity (30), Exception Economics (25), Audit Defensibility (20), Operational Fit (15), and Scale Readiness (10). The decision uses the data captured since day one, compared against the success criteria written before the pilot began. If the pilot met its threshold, the question becomes cost and feasibility of scaling. If it did not, the sunk cost is small enough to walk away.

How do you exit a pilot without sunk cost fallacy?

The sunk cost of a pilot is small and the sunk cost of a half deployment is not. This is the clause that gets resisted most and matters most. After it, sunk cost and demonstration enthusiasm are at their highest, stopping becomes political, and deferral can become permanent. Before the pilot, nobody is attached to the result. Exit criteria must be written before the pilot starts, when no one has invested ego or effort.

Defining explicit exit criteria and choosing a reversible infrastructure stack ensures you can stop a proof of concept cleanly without stranded costs or vendor lock in. When an engineering team connects an experimental large language model to internal databases, builds custom orchestration wrappers, and provisions reserved cloud capacity, walking away feels like discarding sunk capital. In wholesale, reversibility means the pilot does not require custom integrations into order management or ERP systems that would be expensive to remove.

The stop decision is not failure. Validation against baseline: compare the pilot to the manual process over the same periods. For pricing, measure realised margin on the treated categories against a control group of similar untreated categories. For forecasting, compute mean absolute percentage error for both. If the pilot does not beat the current process, stop. A clean stop after 90 days preserves resources for the next attempt and signals that the organisation prioritises evidence over momentum. It is the structure that allows honest evaluation, not the technology, that determines whether a pilot produces a decision or just another extension.

What does a go or no go decision look like in practice?

The decision requires three inputs: quantitative performance against baseline, total cost of ownership at scale, and operational readiness. Click through on AI recommendations against a non personalised baseline, conversion rate, average order value, repeat purchase rate: lift over the control group, not over last year. If the AI improved margin by 80 basis points in the pilot categories, can that result be replicated across the broader catalogue? If order processing time fell by 20 percent, does the team have capacity to handle the exceptions the AI cannot resolve?

According to S&P Global, large enterprises with over 10,000 employees abandoned an average of 2.3 AI initiatives in 2025 alone, with each abandoned initiative carrying an average sunk cost of 7.2 million dollars. The alternative to a go or no go decision at day 90 is not continuation. It is drift into what the industry now calls pilot purgatory: the AI runs indefinitely in a supported but unscaled state, consuming resources without delivering full value. The 90-day framework prevents that outcome by forcing the decision early, when the cost of stopping is still manageable.

Frequently asked questions

What is the most common reason wholesale AI pilots fail to scale?

Most pilots fail because they were never set up to produce a decision. Without baseline metrics captured before the pilot starts, control groups to isolate AI impact, and written success criteria agreed before anyone falls in love with the demo, there is no defensible way to prove value at day 90. The pilot drifts into an extension instead of ending in a scale or stop decision.

How large should the control group be in a wholesale AI pilot?

The control group must be large enough to produce statistically comparable results but small enough to leave volume for the treatment group. Research in e-commerce A/B testing typically uses a 50/50 split, though pilots have run successfully with control groups as small as 20 percent of total volume. The key is that the control must be randomly assigned and truly comparable in customer mix, product category, and time period.

Can a pilot be extended beyond 90 days if results are promising but inconclusive?

An extension is possible, but it must be structured as a new pilot with new success criteria, not a deferral of the original decision. If the 90-day pilot did not produce enough data to decide, the question is why: was the baseline wrong, the scope too narrow, or the metric poorly chosen? Extending without addressing the root cause produces the same inconclusive result at day 120. If the results are promising, the better path is to stop the pilot, document what worked, and launch a scale plan with appropriate investment.

Stories