AI Business Strategy

When an AI agent does the shopping, the seller has to keep the evidence

By Harvey Dhillon, Founder and CEO of Zmartly

Software that buys on a consumer’s behalf is moving from demonstration to live checkout. For small UK sellers on Amazon, Shopify, eBay, Etsy and TikTok Shop, the biggest change is not in the shop window. It is in the records, refunds, tax evidence, pricing and fraud checks behind it. 

Picture an order that arrives at three in the morning. The buyer looked at 40 listings in a few seconds, read your returns policy in full, checked the delivery charge and paid within a budget somebody set last Tuesday. Now picture the same order coming back three weeks later as a dispute. The buyer cannot tell you why it chose you. The person who sent it may not remember sending it at all. 

That is the practical shape of agentic commerce for a small online seller. Most of the commentary on AI shopping agents asks how to get your products found by them. I run an accountancy firm whose clients mostly sell online, so I look at the other end: what happens after the agent presses buy. My short answer is that when a machine does the buying, the burden of proof drifts towards the seller. Sellers who keep good records will find agent orders unremarkable. Sellers who do not will find them expensive. 

This is no longer a forecast 

Three pieces of plumbing arrived within a month of each other in autumn 2025. Google announced its Agent Payments Protocol (AP2) on 16 September 2025. Stripe and OpenAI released the Agentic Commerce Protocol on 29 September 2025, letting ChatGPT users in the United States buy from Etsy sellers inside the chat. Visa unveiled its Trusted Agent Protocol on 14 October 2025 to help merchants tell legitimate agents from malicious bots. 

By 2 July 2026, Visa was announcing live agentic transactions in Europe, with AI agents browsing, selecting and buying from participating merchants, including lastminute.com and Frasers, “acting within consumer-defined parameters”. 

It is still early. The UK’s Competition and Markets Authority (CMA) published a paper on agentic AI and consumers in March 2026. It describes what consumers meet today as mostly “early shopping agents that can search, compare and initiate simple actions with user confirmation”. That makes now the right moment to sort out the back office: before the volume arrives, not after. 

Every buyer becomes a diligent one 

A shopping agent does not turn anyone into a genius buyer. It turns every buyer into a diligent one: a buyer who compares every price, reads every term and never forgets what the listing said. The sellers with most to lose are the ones whose margin quietly depended on customers not checking. 

The agent also holds you to the letter of your listing, while any record of why it chose you sits with the agent’s provider, not with you. Here is how that plays out in five places. 

1. Records: keep the authority with the order 

Google framed the problem plainly when it launched AP2: “today’s payment systems generally assume a human is directly clicking ‘buy’ on a trusted surface”. Its answer is a chain of signed “mandates”, recording the intent the consumer gave the agent, the cart they approved and the payment. Google says the sequence “creates a non-repudiable audit trail” for questions of authorisation. 

For a small seller, the lesson is simple. Whatever your platform or payment provider passes you about the agent and its authority, keep it with the order, for as long as you keep the order. In the UK, HM Revenue and Customs (HMRC) requires value added tax (VAT) records to be kept for at least 6 years, or 10 years under the One Stop Shop scheme, and a customer has up to 6 years to make a claim for goods they bought (5 years in Scotland). Agent data that disappears when a platform tidies its logs is no use in year three. 

The money needs the same discipline. An agent order may settle through a different route from your usual marketplace payout, but it has to land in the same books. UK sole traders and landlords with qualifying income over £50,000 in 2024 to 2025 have had to use Making Tax Digital for Income Tax from 6 April 2026, with the threshold falling to £30,000 from 6 April 2027. A new sales channel has to flow into those digital records from the first order, not be bolted on at the year end. 

2. Refunds and disputes: the listing becomes the contract 

An agent placing the order does not take away the consumer’s rights. The CMA’s position is that “UK consumer law applies whether decisions are made by people or by AI”. For an online sale in the UK, the government’s guidance is that a customer can cancel within 14 days of receiving the item, has another 14 days to send it back, and must be refunded within 14 days of you receiving it. I expect agents to make those rights easier to use, not harder. Software that can buy in seconds can cancel in seconds. 

So, first, your listing stops being marketing copy and becomes a specification. If the agent was told “waterproof, size 10, under £60” and bought your product because the listing said so, a vague description is not a style problem. It is a returns problem. 

Second, take a snapshot. The government’s guidance on distance selling says pre-order information, including “the price, including all taxes”, delivery costs and cancellation terms, must be given in a format “the customer can save for future reference”. The customer’s agent will save it. You should too: the listing text, the price and the terms as they stood at the moment of the order. 

The government’s refunds guidance also says sellers “only have to accept returns from the person who bought the item”. When an agent bought it, that person is the consumer behind the agent, which is exactly the detail Visa says merchants risk losing. Its protocol was built partly for “preserving visibility into the consumer behind the agent”. Keep the customer’s identity tied to the order, not only the agent’s. 

And if you use AI on your own side to handle refunds, the CMA’s guidance for businesses is blunt: “if an AI agent you use does something illegal, you are responsible.” A machine refusing a valid refund to another machine is still you refusing it. 

3. VAT evidence: who is the customer? 

An agent can shop for a household or for a business, and VAT treats those two buyers differently. If you are VAT-registered and sell standard or reduced-rate goods to another VAT-registered business, VAT Notice 700/21 says you must issue a VAT invoice, normally within 30 days. And for goods sent into the UK from abroad in consignments of £135 or less, the marketplace must charge VAT at the point of sale unless the sale is business to business and the customer has given a UK VAT registration number. 

In both cases the tax treatment turns on a piece of data that a person used to type into a box. Now an agent passes it on, or does not. If an agent buys stock for a business and drops the VAT number on the way, the sale looks like a consumer sale. The business then comes back for a VAT invoice it can reclaim on, and somebody has to unpick the order. 

The practical step is to make sure your checkout, or your platform’s, can accept a VAT number from an agent, and then to check it. HMRC’s Check a UK VAT number service confirms whether a number is valid and whose it is, and a VAT-registered business can use it to “prove when you checked”. For an order placed by software, that dated proof is worth having. 

4. Pricing data: two machines and your margin 

Amazon sellers can already let software set the price. Amazon’s own developer documentation says that once a product is linked to an automated pricing rule, “Amazon executes the selected pricing strategy 24 hours a day”, within the minimum and maximum the seller sets. Add a buying agent that checks every competing offer and you have two machines negotiating over your margin while you sleep. 

The CMA has flagged that “algorithmic pricing can also increase the risk of more coordinated market outcomes, even in the absence of explicit communication between businesses”. That is for regulators to watch. The small seller’s more immediate problem is the minimum price. 

What I see ecommerce sellers get wrong is not knowing what profit they actually make. High turnover means cash comes in constantly, and cash flow is easy to mistake for profit. A repricer with a floor set on instinct, facing a buyer that always finds the cheapest offer, turns that mistake into a policy. 

Take a VAT-registered seller with a product priced at £24 including VAT. These figures are illustrative. At the standard rate of 20 per cent, the VAT inside a VAT-inclusive price is one-sixth, so £4, leaving £20. Say the marketplace fee is 15 per cent of the price, £3.60, fulfilment costs £3.20, the product costs £8 landed and advertising takes £1.50. Profit per sale, before overheads: £3.70. 

Now let a repricer chase an agent-friendly price down to £21. The VAT falls to £3.50, leaving £17.50. The fee falls to £3.15 and every other cost stays where it was. Profit per sale: £1.65. A 12.5 per cent price cut has taken 55 per cent of the profit. 

Set the floor from the full unit cost, with VAT and expected returns included, not from a competitor’s price. And keep the price shown at the moment of each agent order. If a dispute turns on what the agent was offered, your pricing history is the evidence. 

5. Fraud: an agent is not a bot, but a bot can pretend to be one 

For years, small sellers have been told to block automated traffic. When Visa launched its protocol, it said merchants face “bot detection systems that can mistakenly block legitimate agentic transactions”, and cited a rise of over 4,700 per cent in AI-driven traffic to US retail websites over a year. The temptation will be to switch protection off and let the agents in. That is the wrong lesson. 

The CMA puts the risk in one sentence: “Weak or fragmented identity systems would raise risks of fraud, error and dispute, undermining confidence in delegation.” The industry’s answer is verification, not openness: signed agent identity through the card networks, and authorisation evidence such as Google’s mandates. In Visa’s European transactions, each purchase was “directly linked to a verified user and their explicit instruction”. 

Then there is the dispute that begins “it wasn’t me, it was my agent”. The answer to that is the authority record from section one. A seller who kept it can show what was approved. A seller who did not is arguing with a shrug. 

On Amazon, eBay, Etsy or TikTok Shop, the marketplace decides which agents get in, and that is still being argued over. In August 2026 a US appeals court lifted an injunction Amazon had won against Perplexity’s Comet browser agent, treating the user rather than Perplexity as the one accessing Amazon, and the case continues. On your own Shopify site the decision is yours, so make it deliberately: accept agents that arrive verified through your payment provider, and treat unverified automated orders like any other high-risk order. 

A framework: the three-question order file 

Everything above comes down to one test. For any order an agent places, could you answer these three questions from your own records, six years from now? 

  1. Who was the buyer? The consumer or business behind the agent and, for VAT, whether they gave a valid VAT number and when you checked it. 
  2. What did they authorise? The agent’s authority as your platform or payment provider passed it, and the listing, price and terms exactly as offered at that moment. 
  3. What can you prove about the money? The payment, the settlement, the fees, the VAT and any refund, landed in the same digital records as every other sale. 

If the answer is yes, an agent order is just an order. If not, close the gap now, while agent orders are a trickle, rather than when they are a meaningful share of your sales. 

The customer who checks everything 

I reconcile my own household accounts every month, the way I would a business, and it is a habit I would recommend to most people. You’ll definitely cancel a lot of subscriptions once you start reconciling your numbers every month. The CMA’s paper imagines agents doing that for consumers at scale, “flagging unused subscriptions” and switching providers on their behalf. 

That is the real shift. Agentic commerce gives every customer the diligence of someone who checks everything, all the time. A seller whose product, price and paperwork survive that customer has little to fear from machine buyers. A seller who relied on nobody checking should start checking first. 

Related Articles

Back to top button