AI in ecommerce: what's real, what's hype, and what to actually implement
Separating signal from noise on AI for ecommerce teams in 2025. We look at where AI genuinely moves revenue and where it's still a distraction.
AI is the most over-claimed topic in commerce right now. Plenty of it is genuinely useful; plenty is a feature looking for a problem. The useful test is simple: does it move revenue, cut cost, or save time on a task you actually do? If a vendor can't answer that in one sentence, it's a distraction.
What's genuinely working
Search & discovery
Semantic and vector search meaningfully outperform keyword search for stores with large or poorly-tagged catalogues. Customers describe what they want in natural language and find it. This is one of the clearest revenue wins available today.
Support deflection
LLM-backed support assistants — grounded in your real policies, order data, and product information — resolve a large share of routine tickets (where is my order, returns, sizing) without a human. The key word is grounded: connected to your systems, not improvising.
Content production
Drafting product descriptions, meta data, and content variations at scale, with human editing. Not a replacement for brand voice — a force multiplier for the volume work that otherwise never gets done.
Merchandising & personalization
Recommendation engines and predictive merchandising have quietly worked for years. The current models are simply better. For stores with enough traffic to train on, the uplift is real and measurable.
What's mostly hype
- Fully autonomous "AI agents" running your store end to end — brittle and unaccountable in practice
- AI-generated lifestyle imagery for premium brands — still uncanny enough to cost trust
- Chatbots bolted on without access to real order or product data — they frustrate more than they help
- Predictive analytics dashboards with no clear action attached to the prediction
How to evaluate an AI feature
Ask three questions before you buy or build: What specific metric is this supposed to move? What data does it need access to, and do we have it clean? And what happens when it's wrong — who's accountable, and what's the fallback? Features that survive those questions are worth piloting. The rest are someone else's roadmap, not yours.
Our position: Adopt AI where it's grounded in your own data and tied to a number you care about — search, support, content, merchandising. Ignore the autonomous-everything pitch for now. The boring applications are the ones paying off.
Want this applied to your store?
We help brands apply exactly this kind of thinking to their actual business — working directly with senior specialists, not account managers.
Commerce engineering, in your inbox.
Occasional, technical, and genuinely useful — engineering perspectives on building faster, better commerce. No fluff.