Someone asked ChatGPT for a pair of merino socks last night, got three options through agentic commerce AI checkout, tapped buy, and paid without ever seeing a website. It is not a 2030 story. It shipped in 2025; it is expanding merchant by merchant, and most New Zealand retailers have no idea whether their products are eligible or invisible.
I am not going to tell you to panic or to rebuild your store. I am going to tell you what actually changed, what breaks first in a typical NZ retail business when an AI agent tries to buy from you, and the specific groundwork worth doing this quarter. Most of it is unglamorous: product data, stock accuracy, returns terms, and whether your back office can answer a question without you.
Key Takeaways
- Agentic commerce AI checkout lets a buyer complete a purchase inside ChatGPT or Perplexity, so the assistant becomes the shopfront instead of your website.
- Instant Checkout went live in ChatGPT in 2025 with Etsy and Shopify merchants, and the protocol behind it was published openly for others to adopt.
- Most NZ retailers are not blocked by payments. They are blocked by product data that is thin, stale, or inconsistent across channels.
- An AI agent cannot buy what it cannot read: identifiers, stock levels, shipping rules and return terms all have to be machine-readable.
- When the agent handles discovery and checkout, you lose the browsing session, so first-party data and post-purchase contact matter far more.
- Expect thinner attribution. Track assistant referrals separately from organic search, or you will misread what is driving revenue.
- The retailers who win are the ones whose back office can answer an order question in seconds without the owner touching it.
- Start with a product feed audit, a stock accuracy check and a written returns policy. That is the groundwork every protocol needs.
What Agentic Commerce AI Checkout Actually Means
Strip the jargon out, and it is simple. For twenty years, the buying flow ran the same way: search, click a link, land on a website, browse, add to cart, check out. The retailer owned the middle of that. Every step happened on your property, with your pixels firing and your merchandising doing the work.
Agentic commerce removes the middle. The buyer talks to an assistant. The assistant searches, compares, picks, and completes the transaction on the buyer’s behalf. The retailer still fulfils the order and still gets paid, but the discovery and the decision now happen somewhere you do not control.
The mechanics, briefly
OpenAI shipped Instant Checkout in ChatGPT in 2025, starting with Etsy sellers and then opening to Shopify merchants, and published the Agentic Commerce Protocol it runs on as an open standard built with Stripe. The design keeps the merchant of record as the merchant. The assistant passes a structured order to your systems, your payment processor takes the money, and your fulfilment process runs as normal. You keep the customer relationship on paper. What you lose is the session.
Perplexity moved on a similar line from a different angle, first with buying inside its shopping answers for paying subscribers, then with an agent inside its browser that can go and complete purchases on retail sites directly. That second version caused a public legal fight with Amazon in late 2025 over whether an agent is allowed to shop on a customer’s behalf inside someone else’s store. Google has been pushing its own agent payments standard alongside both.
Three different approaches, one direction of travel. The purchase is moving into the assistant.
Why the standards fight matters to you
You do not need to pick a winner. You need to be readable by all of them. Each protocol expects roughly the same things: a clean product catalogue, accurate availability, honest shipping and delivery data, and a returns policy written in terms a machine can parse. A retailer who has that is a short integration away from any of these systems. A retailer whose stock counts are a rough guess and whose product titles were written for a human eye is not.
That is the real work, and it is the same work whether the buyer is a person, a bot, or a search engine.
Why Agentic Commerce AI Checkout Hits NZ Retailers Differently
I work with NZ operators every week, and the local version of this problem has its own shape. Three things make it distinct.
We are small, so we are late by default
A New Zealand retailer doing $2M a year does not have a data team. They have a founder who also does the buying, a part-timer on socials, and a bookkeeper. When a new commerce standard lands, the Australian mid-market has a project team on it inside a month. The NZ operator hears about it eight months later from a supplier.
That lag has historically been survivable. It was survivable for mobile, and it was survivable for marketplaces. It is less survivable here, because agentic checkout is winner-take-most at the moment of the answer. There is no page two. The assistant returns two or three options. You are in that set, or you are not mentioned at all.
Our catalogue hygiene is genuinely poor
I have looked inside enough NZ ecommerce stores to say this plainly. Product data in the average local store is a mess. Titles carry supplier codes nobody outside the warehouse understands. Half the range has no barcode identifier. Variant structures were invented on the fly in year two and never cleaned up. Stock counts are directionally correct at best, because the physical shop and the online store both sell from the same shelf and reconcile once a week.
A human buyer forgives all of that. They squint at a blurry photo, they read the description, they ring you. An agent does not forgive it. It filters on structured fields. If your size attribute is stored as free text in the description, you are not in the comparison.
Cross-border economics are already tight
NZ retailers compete against Australian and US sellers who have scale on freight. When an assistant compares three sellers of the same item, and one of them shows a firm delivery window and free returns, guess where it points. The fix is not to fake a better offer. It is to make sure the offer you do have, which is often faster local delivery and a real human who answers the phone, is expressed in the data rather than hidden in your brand story.
Local advantage that only exists in your About page might as well not exist.
The Four Things That Break First
When I audit a retail business for readiness, the failures cluster. It is almost always these four, in this order.
1. The product feed
Your feed is the interface now. Not your homepage, not your theme, not the hero video. If you already run a shopping feed for Google, you have a head start, but most feeds I open are running on defaults from whenever the app was installed.
The specific gaps: missing GTINs or barcodes, inconsistent brand fields, size and colour buried in titles instead of attributes, no material or care data, and images that are either lifestyle-only or cropped inconsistently. Add to that stale pricing where a promotion ran in the store and never synced, and you have a catalogue that reads as untrustworthy to any system evaluating it.
Fixing this is boring and finite. It is a week of work for a 400-SKU range if someone sits down and does it properly. It also improves your paid shopping performance immediately, which is how you justify the week.
2. Stock truth
An agent that sells a customer something you cannot ship is a problem you will hear about loudly, because the complaint lands in the assistant’s ecosystem as much as yours. Availability accuracy is about to become a trust score, whether or not anyone publishes it as one.
If you run a physical shop and a website off one stock pool with a manual reconciliation, that is the thing to fix. Not next year. This is the same discipline that makes your click-and-collect work and your end-of-month count faster, so it pays for itself regardless.
3. Fulfilment and returns as data
“Ships in 1 to 3 days” written on a shipping page is marketing. A delivery promise expressed as structured data, by region, by weight band, with a cut-off time, is something a system can act on. Same for returns: a 30-day window with conditions stated plainly beats a friendly paragraph that requires interpretation.
NZ retailers have a genuine edge here on domestic orders. Auckland to Christchurch overnight beats a three-week wait from offshore every time. But the edge only counts if it is stated in a field, not a sentence.
4. The humans behind the order
This is the one nobody plans for. Agentic checkout compresses the sale and then hands you all of the aftermath at once. Where is my order? Can I change the size? The card was charged twice. It came damaged.
Volume of enquiry per sale does not drop when the sale gets faster. If anything, it rises, because the buyer never read your FAQ. They never saw your site. Their entire relationship with you began and ended in a chat window, so every question comes to you cold.
If those questions currently land in a shared inbox that the owner clears at 9 pm, agentic commerce will make that worse before it makes anything better.
How to Get Ready for Agentic Commerce AI Checkout in 90 Days
Here is the sequence I would run if this were my shop. It assumes no dev team and no budget for a replatform.
Weeks 1 to 2: audit the catalogue
Export your full product list. Every column. Open it and count how many rows are missing a barcode, a brand, a structured size, a structured colour, a weight, and a primary image that shows the product on a plain background. Write the number down. That number is your real readiness score, and it will be worse than you expect.
Then fix the top 20% of SKUs by revenue first. Not all of them. The long tail can wait a quarter. This is the same principle I use for any automation work: do the slice that carries the money, prove it, then continue.
Weeks 3 to 4: get stock honest
Pick one source of truth for availability. Usually that is your point of sale if you have a shop, or your ecommerce platform if you do not. Every other system reads from it. If you currently update stock in two places, kill one of them this month.
Do a full physical count once, reset the numbers, and then measure variance weekly for a month. If variance is under two per cent, you are in good shape. If it is fifteen per cent, no protocol in the world will save you.
Weeks 5 to 8: express your offer as data
Shipping rules by region with real cut-off times. Returns terms with a clear window and conditions. Warranty where it applies. Put these in your platform’s structured fields, not just on a page. If your platform does not have a field for it, most modern ones do, and the app that adds it costs less than a single missed order.
While you are in there, add schema markup to your product pages. It helps with search, and every assistant currently crawling the web reads it.
Weeks 9 to 12: fix the aftermath
This is where most retailers stop, and it is the part that decides whether agentic checkout is good news or a new source of stress.
Map every question a customer asks after buying. There will be fewer than fifteen. Where is my order, can I change it, how do I return it, is it in stock in the shop, do you have it in another size, can I get an invoice with GST on it. For each one, decide who answers it and how fast. Then automate the ones that are genuinely mechanical.
Order status is mechanical. Return initiation is mechanical. Stock lookups are mechanical. A damaged-goods complaint from a repeat customer is not, and should land with a person.
The test is simple. If the owner disappeared for a week, would any of those fifteen questions go unanswered? If yes, that is the gap, and it existed before agentic commerce. This just makes it expensive.
Where the proof sits
I have seen the same pattern outside retail. A finance broker I worked with had 319 contacts in his database that the team had written off as dead. Not a lead generation problem, a follow-up capacity problem. Automated multi-touch reactivation recovered $49,000 from contacts he already owned. The revenue was never missing. The capacity to work it was missing.
Agentic checkout is the same shape. The demand is not the constraint. Your ability to be findable, accurate, and responsive at machine speed is the constraint.
The Part Nobody Mentions: The Business Behind the Checkout
Every article about this focuses on the transaction. The transaction is the easy bit. Payments have been solved for a decade.
What agentic commerce actually does is raise the floor on operational competence. When a buyer browses your site, your presentation smooths over a lot of internal mess. Nice photography and a good brand voice buy you patience. When an agent is evaluating you, there is no patience to buy. It reads your data, compares it to two competitors, and picks. Your operational reality is now your marketing.
That means the useful investment is not a checkout integration. It is the thing underneath: a business where the knowledge is in a system rather than in your head, where the numbers are current rather than reconstructed each Monday, and where the repetitive work happens without someone remembering to do it.
I think of that as two halves. An AI brain, which is the business context and data the system reads before it does anything, and an AI workforce, which is the set of automations that act on it. The brain knows your range, your margins, your delivery zones, your returns terms and your service standards. The workforce answers the order status question at 11 pm, flags the stock variance before it causes an oversell, and chases the abandoned checkout that never made it through.
That is not a new idea I invented for agentic commerce. It is the same system that handles abandoned cart recovery and the same pattern behind any real agentic workflow. Agentic checkout just raises the cost of not having it.
MIT’s research on enterprise AI found that about 95% of initiatives deliver no measurable return, and the pattern in the failures is consistent: they start with a tool instead of a process. Retailers who respond to agentic commerce by bolting on one more app will land in that 95%. The ones who fix their data and their follow-up capacity first will be ready for whichever standard wins.
When This Actually Matters for Your Business
Be honest about your position. If you sell one-off custom work, high-consideration services, or anything where the buyer needs to talk to you before committing, agentic checkout is not coming for your sales this year. Watch it, do the data hygiene because it helps everywhere else, and move on.
If you sell defined products with identifiers, in categories where comparison is normal, and a meaningful slice of your traffic already comes from search, this is your next twelve months. Consumables, apparel, parts, homewares, gifts, pet supplies, anything where the buyer knows what they want and is choosing a seller.
And if you sell through a marketplace as well as direct, pay attention to which channel gets picked up first. Your marketplace listings may become agent-visible before your own store does, which changes your margin mix without you deciding anything.
The middle position, which is where most NZ retailers sit, is to do the groundwork now because it pays regardless. Better product data lifts paid shopping performance. Honest stock reduces refunds. Structured shipping terms cut enquiry volume. Automated post-purchase answers give you back your evenings. None of that is speculative. It all banks value whether or not an assistant ever sends you a single order.
Conclusion
Agentic commerce AI checkout is not a threat to retailers who have their house in order. It is a distribution channel that rewards accuracy and punishes vagueness. The purchase is moving into the assistant, and the assistant only knows what your data tells it.
The work is not glamorous. Clean your feed, make your stock true, write your terms as data, and automate the questions that come after the sale. Do that, and you are ready for whichever protocol wins, and you have a better business in the meantime.
The deeper issue is the one underneath every version of this: a business that runs out of one person’s head cannot respond at machine speed. Fix that, and the rest of it becomes a configuration problem rather than a crisis.
If you want a second opinion on where your business would break under this, grab a 30-minute discovery call. No pitch deck. I will ask what happens after someone buys, and we will find the gap in about ten minutes.
Frequently Asked Questions
What is agentic commerce in simple terms?
Agentic commerce is when an AI assistant completes a purchase for a person rather than just recommending one. The buyer describes what they want, the assistant searches across sellers, compares options, and finalises the order and payment inside the chat. The retailer still fulfils and still gets paid, but the browsing and decision happen outside the retailer’s website.
Can you actually buy things inside ChatGPT right now?
Yes, for eligible merchants and regions. OpenAI launched Instant Checkout in 2025, starting with Etsy sellers and then extending to Shopify merchants, and published the underlying protocol openly. Availability varies by country and category, and the merchant list is still expanding. New Zealand sellers should check eligibility through their ecommerce platform rather than assuming they are already included.
Do I need a developer to get ready for AI checkout?
Usually not for the groundwork. Cleaning your product feed, correcting stock counts, and adding structured shipping and returns terms are all done inside your existing ecommerce platform. A developer becomes useful later if you want custom integration or if your stock lives in a system that does not talk to your store. Start with the data work first.
Will agentic checkout kill my website traffic?
It will reduce the browsing session for transactional purchases, not remove your site. People still visit for brand research, support, and higher-consideration items. The bigger change is attribution. Assistant-originated orders will not look like normal organic search in your analytics, so set up separate tracking early or you will misread which channels are actually producing revenue.
How does this affect GST and invoicing for NZ sellers?
You remain the seller of record under the current protocol designs, so your normal GST obligations and invoicing process apply. The practical issue is that buyers arriving through an assistant often need a GST invoice after the fact and have no account with you. Make sure your system can issue one on request without someone building it by hand each time.
What is the first thing I should fix?
Export your product catalogue and count how many items are missing a barcode, a brand field, a structured size or colour, and a clean primary image. That number tells you your real readiness. Fix the top 20% of products by revenue first, then work down. It takes about a week for a mid-sized range and improves paid shopping results immediately.
Is this relevant if I sell services rather than products?
Less so for checkout, more so for discovery. Assistants are already recommending service providers, and they read the same structured information: what you do, where you operate, your pricing model, and your availability. The follow-up capacity point still applies. If an assistant sends you an enquiry at 9 pm and nobody replies until Tuesday, the referral was wasted.
About Octavius
Titus Mulquiney is the founder of Octavius AI, where he builds AI brains and AI workforces for founder-led businesses stuck running everything out of their own head. Twenty years in marketing, ex-Sony product manager, ex-GM Zeal NZ. Based in Auckland, working with operators across NZ, Australia, and the US. Connect on LinkedIn.