AI in eCommerce: what's actually working now (and what's still hype)

The marketing around AI in eCommerce has outpaced the commercial reality by about two years. Vendors are shipping AI features faster than brands can evaluate them. Agencies are pitching AI implementations without the measurement infrastructure to know whether they worked.

Here's what we're actually seeing generate returns, and what's still in the aspirational category.

What's generating real returns

Search that understands intent

Site search is the highest-ROI application of AI in most Shopify stores right now. The improvement from keyword search to semantic search is substantial and measurable: users searching for "office appropriate dress for summer" finding relevant products instead of zero results, or "gifts under fifty pounds for her" returning curated results rather than a chaos of keyword matches.

The returns are significant because searchers already have high intent. If they're searching, they want to buy. Converting more of them costs nothing on the demand side; it's pure execution.

Solutions like Searchanise, Boost Commerce, and Shopify's native semantic search (rolled out in 2024) have made this accessible without custom ML infrastructure. The gains range from 15–40% on search conversion rate depending on how broken the previous search was. Measurable within weeks.

Product copy at scale

AI-generated product descriptions have moved from novelty to standard practice for brands with large catalogues. The use case isn't replacing copywriters — it's handling the long tail of SKUs that previously got a template description or nothing.

A footwear brand with 800 SKUs used to have well-written copy on the top 100 sellers, template copy on the next 300, and almost nothing on the tail. AI handles the tail competently: accurate product attributes pulled from structured data, brand-voice-consistent prose, SEO-appropriate description length. Copywriter time gets concentrated on hero products where voice and tone matter most.

The caveat: AI-generated copy at scale without editorial review creates risk. Hallucinations in product descriptions (wrong materials, wrong care instructions, wrong sizing claims) are a real problem. The workflow that works is AI draft + structured human review for factual accuracy, not AI publish.

Customer service deflection (for specific query types)

The chatbot use case gets overpromised. "AI customer service" often means a bot that handles simple queries, frustrates customers with complex ones, and creates a worse experience than a well-structured FAQ page.

What actually works: AI handling the narrow, high-volume, factual query types. Order status. Delivery windows. Return policy specifics. Size guide lookups. These queries have structured, correct answers. AI retrieval against your own data produces reliable responses. The deflection rate on this narrow category is typically 60–80%, which is meaningful if those queries represent a large chunk of your support volume — and for most DTC brands, they do.

What doesn't work: AI handling returns disputes, complaints, product recommendations requiring judgment, or any query where the correct answer requires nuance. These need humans, and routing them incorrectly through AI first makes the eventual human handoff harder.

Inventory and demand forecasting

For brands with sufficient data history, AI-assisted demand forecasting is producing genuine planning improvements. The category where this is most visible: seasonal businesses and brands running limited-edition drops.

The traditional approach — sales history + gut feel + buyer experience — leaves money in stockouts and working capital tied up in slow-moving inventory. ML-assisted forecasting on top of two to three years of sales data, promotional calendars, and external signals (search trend data, social velocity) produces meaningfully better input for buying decisions.

This requires data infrastructure investment. If your sales data isn't clean, tagged, and accessible for modelling, AI forecasting is premature. Fix the data layer first.

What's still mostly hype

Full personalisation at the individual level

The pitch: every customer sees a different homepage, different product ordering, different pricing, different messaging based on their individual profile. The reality: this requires data depth and infrastructure that most mid-sized brands don't have, and the lift over good segment-based personalisation is smaller than claimed.

Shopify Audiences and segment-based merchandising (showing different collections or landing pages to different acquisition segments) delivers most of the commercial benefit at a fraction of the complexity. Start there before buying into 1:1 personalisation infrastructure.

AI agents managing your eCommerce operation

The agent use case — an AI that autonomously manages replenishment, responds to supplier emails, adjusts pricing based on competitor data, and approves ad spend — is real in controlled demos. In production eCommerce, the error rate on unsupervised decisions is too high for most operators.

Where agents are genuinely useful today: supervised workflows with human approval gates. AI drafts a replenishment order based on forecast data; a buyer approves or adjusts it. AI identifies underperforming SKUs and suggests promotional actions; a merchandiser confirms and implements. The "agent" is reducing the cognitive load on a human decision-maker, not replacing them.

The fully autonomous version is coming. It's not ready for deployment on operations where errors cost real money.

Predictive lifetime value at small data volumes

LTV modelling gets pitched to brands of every size. It requires enough transaction history to be meaningful: typically 3+ years of data, sufficient order frequency to identify patterns, and a customer base large enough to segment. Below that threshold, the model is overfitting to noise. The predictions aren't reliable, and making inventory or acquisition decisions based on them is worse than not having the model at all.

The practical starting point

If you're deciding where to invest in AI for an eCommerce operation, the priority order that actually generates returns:

1. Search quality — high ROI, measurable quickly, applies to every store
2. Customer service deflection on factual queries — clear ROI, reduces support cost
3. Product copy at scale — value proportional to catalogue size and current copy quality
4. Demand forecasting — valuable if you have the data and the buying scale to act on it

Everything else requires you to have done those well first.