How to Use AI to Run an Ecommerce Store

US shoppers are expected to return $849.9 billion of merchandise in 2025 — 15.8% of all retail sales. Online, the rate is much worse: 19.3% of ecommerce sales come back. Those figures are from the National Retail Federation's 2025 Retail Returns Landscape.

Roughly one in five online orders returning is not a customer service problem. It is a margin problem that happens to arrive through customer service, and it is the clearest place where AI earns its keep in ecommerce — not by writing more product copy, but by finding out why the thing came back.

Diagnose returns before you rewrite anything

The instinct when a SKU has a bad return rate is to rewrite the product page. Often that's exactly wrong, and you can only tell by reading what customers actually said.

Paste your return reasons and reviews, and force the split:

Here are the return reasons and review excerpts for this SKU. Sort the causes into three buckets: (a) the page created a wrong expectation, (b) the product itself has a defect or fit problem, (c) the customer ordered multiples intending to return some. For each, quote the evidence. Don't propose fixes yet.

The buckets have completely different owners. A wrong expectation is a photography, sizing-chart or copy fix and you can ship it this afternoon. A defect is a supplier conversation. Bracketing — ordering three sizes to keep one — isn't a fault at all; it's a demand-planning input, and fixing it with a stricter policy costs you the sale.

Getting this wrong is expensive in a specific way: you rewrite the page, the return rate doesn't move, and you conclude the copy wasn't the problem when the real answer is that you fixed the wrong bucket.

Structured data is now the product page

The old job was writing a description a human would enjoy. The new job is emitting attributes a machine can parse — because a growing share of discovery runs through marketplace algorithms, shopping feeds and AI assistants answering waterproof hiking boots for wide feet under $150.

A model can't answer that from your evocative paragraph about the mountains. It answers from attributes: width fitting, waterproof rating, upper material, weight, price. If those aren't explicit and structured, you are invisible to the query regardless of how well the page reads.

So use AI in the direction that matters: extraction, not generation.

From this supplier spec sheet and my description, produce a complete structured attribute set: material, dimensions, weight, care, compatibility, sizing, country of origin. Mark anything you cannot find as [MISSING]. Do not infer or estimate any value.

That final instruction is load-bearing. Left alone, a model will fill a gap with a plausible number, and a plausible-but-wrong material or weight in a product feed is a returns problem, a compliance problem, and on some marketplaces a suspension problem. [MISSING] is a task for a buyer. An invented value is a liability that looks complete.

This is why the ecommerce pack treats product data, listings and returns as one connected job rather than as copywriting.

Returns policy: the counterintuitive part

The NRF data makes the trade-off unusually concrete. 82% of shoppers say free returns are a major purchase consideration — up from 76% the year before — and 71% say they're less likely to shop with a retailer again after a poor returns experience. Meanwhile about 9% of returns are fraudulent.

Which means the obvious lever — tighten the policy to stop abuse — is usually the wrong one. You'd be taxing the 91% to deter the 9%, and losing repeat purchases at the same time.

Use a model to write for both readers at once:

Draft a returns policy that a good-faith customer finds reassuring in the first two sentences, and that closes the specific loopholes below further down: [worn-and-returned, missing accessories, serial-number swaps]. Plain English, no legalese. Flag any clause that needs a lawyer or varies by jurisdiction.

Then have it list what your policy doesn't say. Ambiguity is where disputes start, and gaps are much easier for a model to spot than for the person who wrote the policy.

Inventory: bring the numbers, ask for the reasoning

Reorder quantities and peak-season plans are the highest-stakes decisions in a small store, and they're the ones where a model's arithmetic is least trustworthy. Multi-SKU calculations across lead times and sell-through rates produce confident tables with errors buried in them.

The reliable division: you supply sales history, lead times, current stock and cash constraint; the model surfaces the decision structure — which SKUs are stockout risks, where the cash is trapped, what the reorder trigger should be, what happens if the container is two weeks late. Then verify every number that leads to a purchase order.

Ask it to state its assumptions explicitly before the recommendation. Most bad reorder advice traces to an unstated assumption about demand continuing flat.

What not to hand over

Pricing against competitors. A model doesn't know their costs, their inventory position, or their strategy, and will happily rationalise a race to the bottom.

Anything legally binding without review. Policies, warranty terms, and compliance claims vary by jurisdiction and get flagged as fabrication risk for good reason.

Review responses at scale, unread. A generic apology under a specific complaint is worse than silence — visibly so, to every future shopper reading the page.

The pattern across all of it: AI is strongest where you have data and need structure — reasons sorted into causes, specs turned into attributes, history turned into a decision. It is weakest where it has to supply facts about your business it was never given. Feed it your numbers, make it mark its gaps, and check the arithmetic before it becomes a purchase order.

---

Source: National Retail Federation, 2025 Retail Returns Landscape (October 2025).