The obvious answer is usually the wrong one
Most e-commerce AI conversations start with personalised recommendations. That is a mature, well-served problem where you are competing with platform features, and it is rarely where a mid-sized retailer’s money is being lost.
1. Product data, which is always a mess
Descriptions, attributes, and categorisation supplied inconsistently by suppliers. Normalising that improves search, filtering, and how AI search describes what you sell — and it is exactly the bulk, verifiable work models handle well.
Your catalogue data quality decides your on-site search, your filters, and what a model says you sell.
2. Support volume that is mostly the same question
Where is my order, can I return this, does it fit. High volume, well documented, and safe to draft for a human — the classic first three jobs with unusually clean economics.
3. Returns, where the reasons are free research
Return reasons written in free text are a direct signal about product descriptions, sizing and photography. Reading them at scale finds the listings costing you money — the exhaust nobody analyses.
Be careful with pricing automation
Automated repricing is fast to build and capable of destroying margin or triggering competitor responses. It belongs in the category of irreversible actions a model should prepare, not commit.
Measure against baseline, not against hope
Return rate, contact rate per order, and on-site search success are all measurable before and after. E-commerce has better instrumentation than most sectors — use it rather than accepting a vendor’s uplift claim.