How agentic commerce works
In agentic commerce the buyer states an outcome — reorder the filter I bought last winter, find a small desk that arrives inside a week — and software does the shopping. The agent reads product pages or a merchant feed, compares what it finds against the instruction, chooses, and then either hands a filled basket back to the person or completes the payment itself.
The hard part is not the shopping, it is the paying. An agent has to prove it is spending on behalf of a real person who agreed to a real amount. OpenAI and Google have each published proposals for how that authorisation could travel with the request, and the card networks are working on credentials issued to agents rather than to people. None of it is settled, and no single approach has won.
Why agentic commerce matters
It moves the point of comparison off your website. If an agent shortlists three shops before a human sees anything, your product page never gets its chance to persuade — the machine-readable version of your catalogue persuades instead. That puts real weight on the product feed, on structured data, and on plain facts such as stock, delivery time and the total price including shipping.
It also arrives unevenly by market. In Nepal most checkouts still finish on eSewa, Khalti, a bank transfer or cash on delivery, and an agent cannot drive any of those on its own. Agentic buying will reach Nepali merchants through card payments and cross-border selling long before it reaches a local shop.
Where agentic commerce goes wrong
The most expensive mistake is treating it as a channel to chase. It is early, the standards are unfinished, and a tool that promises agent-ready checkout today is selling somebody’s guess. Nothing about it justifies breaking a checkout that currently converts.
The second is data that disagrees with itself. If your feed says one price and your page says another, an agent reads the mismatch as a reason to skip you rather than a rounding difference to forgive. The same applies to stale stock levels and delivery promises that only appear at the final step.
The third is what happens afterwards. An automated purchase still has to be reversible by a human, and most return, dispute and fraud processes were written on the assumption that a person clicked the button.
What to do about it
Do the unglamorous work that pays off either way. Keep price, availability, shipping cost and return terms accurate and identical across your feed, your page text and your product schema. Make guest checkout genuinely work. Never make someone ask in a chat window to learn what something costs.
Then watch rather than build. Look for unfamiliar user agents and data-centre traffic in your server logs, check whether AI assistants quote your products correctly when asked, and keep your terms of sale clear about who is deemed to have authorised a purchase. That is enough preparation until the standards firm up.