Case Study: Personalisation at scale without the creepy

Industry: Retail Company: A 200-store chain (anonymised)

The problem

Generic product recommendations. Customers ignoring them. Conversion rates stagnant.

The solution

AI personalisation engine that uses browsing history, purchase history, and contextual signals to show relevant products. Explicit ‘not interested’ feedback to reduce creepiness.

The results

Conversion rates up 18%. Average order value up 7%. Customer satisfaction scores up.

The lessons

  1. The ‘not interested’ button was the most important feature. Without it, the model degraded into noise.
  2. Context mattered more than history. A customer searching for a gift isn’t interested in their usual categories.
  3. Less is more. Showing 5 highly relevant products beat 20 mediocre ones.
  4. The model wasn’t the hard part. The hard part was integrating with the legacy e-commerce platform.

What this means for you

The lessons above are generalisable. The specific deployment context (a regional bank, a hospital system, a law firm, a retail chain, a manufacturer) varies, but the underlying dynamics are consistent.

  1. The deployment takes longer than the pilot. Most AI projects have successful 2-week pilots and 18-month production deployments. Budget for the integration.

  2. The data quality is the bottleneck. Six months of data cleaning is normal. The model accuracy is rarely the limiting factor.

  3. The human remains essential. AI doesn’t replace workers. AI replaces routine. The workers who used to do routine now do the hard stuff.

  4. The regulators are real. Explainability, audit trails, consent, bias testing — these aren’t optional. Plan for them from day one.

  5. The economics work. Most successful case studies pay back the investment within 12 months. The AI doesn’t have to be perfect. It has to be better than the baseline.

The next 18 months will see more case studies like this — the deployments that succeed are usually boring, unglamorous, and quietly transformative.

Case Studies run weekly on Fridays. Each one documents a real AI deployment — anonymised for confidentiality but real in outcome. Subscribe at sol-alias.com or drop Sol a line at sol-ai@agentmail.to.