AI in retail and ecommerce, what actually moves the needle
Where artificial intelligence produces measurable results in retail and online commerce: assortment, pricing, content, service and replenishment.
Retail is one of the industries with the most announced AI cases and the widest gap between what's announced and what actually changes the numbers. These are the applications where impact is verifiable.
1. Product content
The least flashy case and the best return for most online retailers. Catalogues with inconsistent titles, incomplete descriptions and missing attributes lose sales on two fronts: customers can't find products in internal search, and search engines don't rank them.
Generating and normalizing titles, descriptions and attributes from available information, consistently across the catalogue, has a direct effect on conversion. And it requires no complex integration.
2. Internal search
A high share of searches on commerce sites returns poor results because of wording differences: the customer searches with different words than the catalogue uses. Semantic search solves that without anyone maintaining synonym lists.
It's one of the interventions with the best effort-to-conversion ratio available.
3. Demand forecasting and replenishment
The classic case, and the one that moves the most money in retailers with physical inventory. A forecast by product and location incorporating seasonality, promotions and events reduces both stockouts and tied-up inventory.
The requirement is sufficient history and clean master data. Without those, the project stalls before it starts.
4. Service and post-purchase
Enquiries about order status, exchanges, returns and availability make up most service volume and are highly repetitive. This is where automation reduces cost most clearly, provided there's a visible route to a human.
5. Price management
Analysing elasticity, competition and turnover to suggest prices has value, but needs more care than the cases above. Three rules:
- Explicit variation limits by category.
- Excluded products: the ones that define the brand's price perception.
- Bounded frequency: changing the same item's price several times a week erodes trust.
Optimization without business rules works in the spreadsheet and fails in the market.
6. Personalization
Recommendations and communications tuned to customer behaviour. Works well with large catalogues and high traffic. With small catalogues the gain is marginal and investment is better spent on content and service.
What doesn't work as promised
Individual purchase prediction. Knowing exactly what a specific person will buy is less reliable than promised, and actions based on that prediction usually underperform good segmentation.
Fully automated customer service. Returns and complaints carry an emotional component; fully automating them costs more in reputation than it saves in service.
What to measure
- Conversion by channel and category, against an equivalent period.
- No-result search rate: the most direct indicator of catalogue quality.
- Stockouts and days of inventory.
- Service cost per order.
- Return rate, which reveals whether generated descriptions are accurate or inflating expectations.
That last indicator is the quality control for point 1: if descriptions lift conversion but raise returns, the benefit is illusory.
Frequently asked questions
Which AI application should an ecommerce implement first?
Improving product content: consistent titles, descriptions and attributes. It affects internal search, organic ranking and conversion directly, and requires no complex integration.
Is personalization worth it for small catalogues?
Below a few hundred products, the gain from a recommendation engine is marginal. Prioritize content quality, internal search and customer service instead.
How do you run dynamic pricing without damaging brand perception?
By setting clear variation limits, excluding sensitive products and avoiding frequent changes on the same item. Optimization without business rules erodes customer trust.