A customer in Tai Po needs a replacement filter for her kitchen tap. In 2023 she would have typed the brand into Google and clicked through three shops. In 2026 she types one sentence into an AI assistant and asks it to find the part, confirm it fits, and tell her who has it in stock today. She never opens a browser tab.
Your shop is either inside that answer, or it is not. That is agentic commerce, and it is worth understanding before you decide whether it deserves any of your budget this year.
What is agentic commerce?
Agentic commerce is shopping carried out by software on a person's behalf. The customer states what they want and sets limits. An AI agent then searches, compares, checks price and availability, and either recommends one option or completes the purchase. The buying decision moves out of a browser and into an assistant.
The useful way to picture it is a personal shopper who works for the customer, not for you. This shopper reads every product page in your category in about a second, ignores your shopfront design entirely, and cares about exactly five things: what the item is, whether it fits, what it costs, whether it is in stock, and what happens if it is wrong.
Nothing about that list is new. What is new is that a machine is now doing the comparison, and machines do not squint at a blurry photo and guess.
How does an AI shopping agent decide what to recommend?
An agent reads structured product data, not page design. It looks for a clear product name, a price, a stock status, shipping terms, a return policy and comparable specifications. Where those facts are missing or inconsistent, the agent has nothing to place in its comparison, and it usually moves on to a shop that published them.
Think of it as a tender document. The agent is filling in a table with one row per shop. Blank cells do not get a sympathy vote. They get skipped.
The four things an agent needs from your listing
--- Identity. The exact product name, brand, model number and category, written the way a customer would say it rather than the way your supplier codes it.
--- Specifications. Dimensions, material, capacity, compatibility, power rating. These are the fields an agent uses to answer "does it fit".
--- Commercial terms. Price including whether tax and delivery are inside or outside it, stock status, and delivery time to a Hong Kong address.
--- Risk terms. Return window, warranty, and who pays return postage. Agents weight these heavily because their user asked them to avoid bad outcomes.
Did AI checkout actually work in its first year?
Not yet. OpenAI launched Instant Checkout inside ChatGPT on 29 September 2025 and withdrew it on 4 March 2026, roughly five months later. By February 2026, only around 30 Shopify merchants were live on it. Walmart measured in-chat checkout converting about three times worse than a click through to walmart.com.
The reasons reported at the time were unglamorous and instructive. Sales tax collection was not built. Fraud prevention was not built. Real-time inventory could not be kept in sync across merchants at scale. As retail analyst Jason Goldberg wrote in Forbes in March 2026, the retreat said more about operational plumbing than about consumer appetite.
OpenAI's own framing was that Instant Checkout was moving to Apps, routing transactions through partners such as Instacart, Target and Booking.com instead of completing them in the chat window.
The shape that survived the first year is therefore simple enough to put on a sticky note: discover in AI, buy on site. Assistants are getting good at finding and comparing. They are not, in 2026, a reliable place to take money.
That matters for your budget. It means the work worth doing this year is the work that makes you findable and comparable, not a rush to plug into a payment protocol.
What are UCP, ACP and AP2 in plain language?
Three competing open standards now govern how AI agents shop and pay. UCP, from Google and Shopify, covers discovery and cart. ACP, from OpenAI and Stripe, handles checkout execution inside a chat. AP2, started by Google and now governed by the FIDO Alliance, proves who authorised a payment. A single purchase can touch all three.
Google and Shopify unveiled the Universal Commerce Protocol at the National Retail Federation conference in January 2026. From 17 June 2026, any developer can register an agent profile in Shopify's Developer Dashboard and call its public endpoint without an approval gate, which is why UCP has the widest reach among small merchants: they inherited it.
Here is the part most owners get wrong. You are almost certainly not going to implement a protocol. Your platform will, and it will arrive as a feature update you did not ask for. The standards war is a fight between Google, OpenAI, Stripe and Shopify, and your role in it is to have clean data on the day your platform switches something on.
The one question worth asking your platform or web vendor is this: which agent standards do you support today, and what, if anything, do I have to do about it? A vendor who cannot answer that in a sentence is telling you something.
What does this mean for a Hong Kong shop right now?
Referral traffic from mainstream generative AI platforms accounted for roughly 2.5% of traffic in SHOPLINE's 2026 Hong Kong e-commerce whitepaper, after rising about 485% year on year in December 2025. That is a small share on a steep curve. Meanwhile the 2026 Adyen Index found 94% of Hong Kong merchants familiar with agentic commerce and 52% planning to integrate it into their revenue plans.
The same Adyen research recorded what is holding merchants back: 43% worry about losing the direct customer relationship or brand control, 38% cite data privacy and security, and 35% point to the difficulty of integrating AI with existing systems.
A worked example: a six-person homeware shop in Tsuen Wan
The shop lists about 1,200 items online. Every listing has a name and a photo. Roughly 40% have no dimensions, no material and no stock status, because those fields were never mandatory when the catalogue was built.
Now imagine the request an agent receives: find a 28cm stainless steel steamer that fits a standard Hong Kong gas hob, in stock, under HK$400. The agent can only compare shops that published dimensions, material, price and stock. The Tsuen Wan shop stocks exactly the right steamer and is invisible for that question.
Filling three missing fields across roughly 480 incomplete listings, at 40 seconds each, is about 5.5 hours of work. Doing it across all 1,200 listings is around 13 hours, roughly one week of a part-timer's afternoons. That is the entire entry ticket, and it costs less than a fortnight of paid search.
The reason this is a sensible spend even if agents never take off is that the same fields help ordinary search results, comparison sites, and the human customer who was about to message you at 11pm asking whether it fits.
Four things people get wrong about agentic commerce
"It is the same as the chatbot on my website." It is not. A site chatbot serves people who already found you. Agentic commerce decides whether they ever arrive. One is service, the other is distribution.
"I have to join a protocol." Almost no small merchant implements ACP or UCP directly. Your platform does. Spending on protocol integration before your product data is complete is paying for a faster car with no wheels.
"AI shopping will replace my website." The first year of evidence pointed the other way. Checkout retreated to merchant sites precisely because tax, fraud and inventory live there. Your site got more important, not less.
"2.5% is too small to bother with." The share is small and the honest answer is that it may stay small for a while. The reason to act anyway is that the fix is not agent-specific. Complete, accurate product data is the cheapest thing you can do that improves search, ads, marketplaces and agents at the same time.
Frequently asked questions about agentic commerce
Do I need a Shopify store for any of this?
No. Shopify happens to be furthest along because it co-authored UCP, but the underlying requirement is platform-neutral: publish complete, machine-readable product facts. A WooCommerce, SHOPLINE or custom site with clean data will be read; a Shopify store with empty specification fields will not.
Will AI agents expose my prices to competitors?
Your published prices are already public. What changes is the speed of comparison, which pressures shops competing only on price and helps shops that can also state fit, stock and returns clearly. If price is your only differentiator, agents will find that out faster than shoppers did.
How do I know whether an assistant already recommends my shop?
Ask it. Write down the three sentences a real customer would type, run each one in three different assistants, and record what comes back. Repeat monthly. It takes twenty minutes and it is the only measurement most small merchants need in 2026.
Is this only relevant to e-commerce?
No. Service businesses face the same mechanism. An assistant asked to find an aircon technician who covers Kowloon, works weekends and gives fixed quotes is doing exactly the same comparison, using whatever facts you published about coverage, hours and pricing.
What is the single first step this month?
Pick your twenty best-selling items and complete every field a customer would ask about before buying. Twenty listings is one afternoon. If nothing changes, you have lost an afternoon. If something does, you will have found out cheaply.
The takeaway
Agentic commerce in 2026 is not a checkout revolution. It is a discovery shift, and it rewards the least glamorous asset a small shop owns: accurate, complete, boring product information.
The first year taught the industry that machines can compare far better than they can transact. So the sensible position for a Hong Kong SME is neither panic nor dismissal. Publish facts an agent can use, measure whether assistants mention you, and let the protocol war resolve itself without your money in it.
Technology is only useful when it arrives with someone who can explain it plainly. We understand AI. UD stands with you.
Further reading: What Is llms.txt? And Does It Actually Work Yet? and AEO or SEO: Where Should a Hong Kong SME Spend First?
Reviewed by the UD AI team.
Not sure where your shop stands?
Before you spend anything on agents, it helps to know what AI assistants can already see about your business. UD's free AI Ready Check gives you a starting picture in a few minutes, and if you want to go further, we will walk you through it step by step, from fixing your product data to measuring whether assistants start mentioning you.
