Field Notes26 July 20268 min read
What NZ e-commerce operators are actually doing with AI
Product descriptions, after-hours service, email personalisation: where AI is earning real ROI in NZ online retail.
Product descriptions, after-hours service, email personalisation: where AI is earning real ROI in NZ online retail.
The NZ e-commerce sector sits in an unusual structural position. Domestic operators compete against international platforms with local warehousing, freight economics that bear no resemblance to what a Christchurch small business pays to courier overnight to Wellington, and marketing budgets that dwarf the entire NZ SME segment combined. Competing on price is often structurally impossible. The businesses gaining ground are winning on speed, service quality, and content clarity, and AI tools are starting to make a measurable difference across all three.
This is not a product comparison. It is a report on what NZ online retailers have actually deployed, what the numbers looked like after six months, and where early enthusiasm quietly faded.
The most common entry point for NZ e-commerce businesses using AI is product copy. It is also, typically, the use case with the clearest before-and-after numbers.
A Christchurch homewares retailer we worked with had around 2,200 active SKUs across three categories. Their descriptions had been written in batches over several years by different people, had inconsistent structure, variable keyword coverage, and an average length of 85 words. The SEO implications of that inconsistency were obvious. The AI search implications were less obvious but equally significant. A product page that doesn’t clearly describe what the product is, who it’s for, and what problem it solves doesn’t give a language model enough material to confidently recommend it when someone asks Perplexity or ChatGPT to suggest options in the category.
The project: audit existing copy, establish a brief per product category covering what the customer needs to know, what differentiates the product, and what the length floor should be, then generate refreshed descriptions using a language model with that brief as context. A staff member reviewed each batch for factual accuracy and adjusted anything that described a product incorrectly.
Time investment before: an average of 12 minutes per SKU when written manually. After: around 90 seconds per SKU including review time. Across 2,200 SKUs, that is a difference of roughly 400 hours of writing and editing time. The project ran over six weeks with part-time staff involvement. Average description length after the refresh: 210 words. Average time-on-page increased 18% across the refreshed category pages within three months.
The consistent caveat with AI-generated product copy is that it requires good inputs to produce useful outputs. A furniture retailer who feeds a language model “oak dining table, 1800mm x 900mm, seats 8” will get generic output. A retailer who includes the supplier’s material specifications, notes from the buyer about which customers are drawn to this piece and why, notes on the actual grain pattern and finish, and the key phrases customers use in search, gets copy that sounds like it was written by someone who had visited the showroom.
The work is not in the generation. It is in building the brief template that reliably produces output worth publishing. That is a one-time investment that pays back at scale. The operators who have not invested in it are generating content that is technically correct and aesthetically neutral, which is, for products where the purchase is partly emotional, a worse outcome than saying less.
NZ e-commerce customer service queries follow a predictable distribution: shipping ETAs, order status, return requests, sizing or compatibility questions before purchase. These are high-volume, low-complexity interactions that do not require a human decision, but they do require a timely response. And timely, for a buyer in the middle of deciding, means minutes, not hours.
The after-hours problem for NZ e-commerce is sharper than it is for counterparts in larger markets. A significant portion of NZ online shopping happens between 8pm and midnight, outside any staffed service window. Queries that arrive then often sit until the next morning. For a purchase decision already in progress, that wait is frequently the difference between completion and abandonment.
An Auckland apparel retailer implemented an AI assistant on their product and checkout pages to handle pre-purchase queries outside business hours. The assistant had access to product inventory data, sizing guides, and shipping policy, and was configured to escalate anything requiring a human decision or order-specific data.
After three months: 64% of after-hours queries resolved without staff involvement. Average response time dropped from 14 hours to under two minutes for handled queries. The escalation rate was 22%, mostly order-status requests that required access to the fulfilment system, which hadn’t been connected in the initial build.
The design decision that made this work was scope constraint. The assistant was explicitly unable to process returns, make pricing exceptions, or handle complaints. Those paths escalated immediately. The narrow scope was not a limitation; it was the reason the resolution rate was high enough to be useful. An AI automation layer that tries to handle everything handles nothing well.
Most NZ e-commerce platforms include abandoned cart sequences by default. The AI layer adding value is not in those templates. It is in the personalisation of follow-up content based on what the customer actually viewed and when, rather than demographic assumptions about who they are.
A Wellington outdoor and camping retailer worked with us on a post-purchase email sequence that used the customer’s first order to infer what content would be relevant over the following 90 days. Someone who bought a sleeping bag in May received content about gear maintenance in June, layering systems in July, and a targeted product recommendation sequence in August before the main tramping season. The personalisation used purchase data and browse history rather than age brackets or assumed income signals.
Open rate on the AI-personalised sequence: 38%. Their previous broadcast newsletter averaged 22%. The difference was not the subject lines. It was that the content was relevant to what the customer had already indicated they cared about. That 16-point difference, across a list of 11,000 subscribers, represents a significant change in the commercial return on the same send volume.
There is a direct connection between product page content quality and AI search optimisation performance that most NZ e-commerce operators haven’t yet considered. When someone asks an AI search tool to recommend a specific type of product in NZ, the tools generating those answers are drawing on indexed product and category page content. A product page with detailed specifications, use-case descriptions, and comparison content gives a language model material to work with. A product page with a 60-word manufacturer description does not.
The operators who have refreshed their product content for SEO reasons are, incidentally, improving their AI search visibility at the same time. The content requirements overlap significantly. Clear, specific, structured copy that answers the questions a buyer would ask is what both Google and a language model prefer. The businesses investing in content infrastructure are compounding across both surfaces simultaneously, while the ones who aren’t are losing ground on both.
The categories where AI assistance is least effective in NZ e-commerce are consistent across the operators we have worked with.
Products where the purchase decision is primarily emotional or highly taste-dependent, art, handmade ceramics, jewellery, specialty food, don’t benefit from generated copy in the same way commodity products do. The output is typically correct and neutral, which is a worse outcome than copy written by someone who understood what made the product worth making in the first place. For those categories, AI works better as a drafting aid than as a replacement for a writer who actually understands the product.
The second failure pattern is automation at the customer service layer that removes the signals a small NZ business trades on. Customers who buy from a Tauranga ceramicist or a Hawke’s Bay food producer are often buying partly because it is a specific person, a specific place, something with a face behind it. A customer service layer that reads like a corporate ticketing system is working against the brand rather than for it. The test is straightforward: if the communication could have come from any e-commerce business anywhere, it probably shouldn’t have come from yours.
The third failure pattern is the AI-generated product catalogue that skips the review step to move faster. Without a human reviewing for accuracy, errors propagate at scale. A product described as available in a colourway it doesn’t come in, or with dimensions that are wrong by 5%, damages customer trust in ways that a small, careful rollout avoids entirely. Speed without review is not a use of AI; it is a liability.
The NZ e-commerce businesses best positioned heading into 2027 are the ones building content and service infrastructure now rather than running one-off campaigns. Comprehensive product descriptions. Structured FAQ content. Post-purchase communication sequences built on actual customer behaviour. The businesses investing in that layer are compounding across Google visibility, AI search citations, and on-page conversion at the same time.
The next meaningful capability shift for NZ e-commerce AI is the integration of real-time inventory and fulfilment data into customer-facing service layers, the piece that required manual escalation in the Auckland example above. When that connection is reliable enough to deploy without the risk of surfacing incorrect availability information, the resolution rate on AI-handled customer queries will climb significantly. The operators who have already built the first layer will extend to that one quickly. The ones who haven’t will be starting from scratch while their competitors are already iterating on their second version.
The gains in NZ manufacturing aren't on the factory floor. They're in the admin layer that surrounds it.
Three areas where NZ construction firms are finding real returns from AI, and one where most are still leaving time on the table.
Local search, email, and social: where AI is changing the economics of NZ physical retail.