We looked recently at AI shopping assistants and what they mean for e-commerce brands, and the reassuring finding there was that the assistant mostly hands the shopper back to you: it advises, you transact, and the visitor who arrives on your product page is unusually warm. That's still broadly how the money moves. But it rests on one word that is quietly changing underneath the industry — assistant.
An assistant advises. An agent acts: it takes a goal and a set of constraints, evaluates options, and completes the purchase itself. That's not a semantic distinction, it's a commercial one. An assistant still delivers a human being to your page, where your photography, your copy and your layout get to do their job. An agent may never load the page at all.
The rails got built while everyone argued about chatbots
The reason this stops being hypothetical is infrastructure. Since late 2025 a set of open protocols has emerged specifically to let AI systems transact with merchants — OpenAI and Stripe's agentic commerce protocol arrived first, Google and Shopify introduced a competing universal standard at the January 2026 NRF conference, and the Model Context Protocol has become the common way for an AI system to query a retailer's live inventory and pricing directly rather than scraping a rendered web page.
The payment networks moved in behind them. Through early 2026, Visa commercialised a trusted-agent protocol, Mastercard ran live agent-initiated transactions, and American Express shipped developer tooling for agent purchases. Salesforce's retailer research has found a large majority of retailers now expect AI agents to be essential to their business, and Adobe's holiday data indicated AI-referred visitors complete purchases at a notably higher rate than those arriving from traditional search.
None of that means your customers are all delegating their shopping tomorrow. It means the plumbing exists, the card networks have blessed it, and the cost of being unprepared has stopped being theoretical.
Your conversion toolkit assumes a human is looking
Here's the part that should genuinely unsettle a retail team. Almost everything you have spent a decade optimising is aimed at a nervous system that may not be present.
Hero photography, art direction, the carefully sequenced landing page, the urgency banner, the "only three left" badge, the exit-intent offer, the carefully worded brand story — none of it registers with a machine evaluating your product against three competitors on the attributes its owner specified. PwC put the mechanics of this bluntly in its agentic commerce work: if a shopper tells their agent to optimise for price, a $98 listing beats a $100 one, every single time, and no amount of craft on the $100 page changes the arithmetic.
The shift in one line For twenty years, e-commerce has competed on being persuasive. Agentic commerce competes on being selectable — clear, certain, verifiable, and easy to execute against. Those are not the same skill.
| What you optimise for humans | Does an agent register it? | What it weighs instead |
|---|---|---|
| Photography and art direction | Essentially no | Structured attributes in your feed |
| Urgency and scarcity cues | No | Real availability and delivery date |
| Brand story and persuasive copy | Barely | Spec completeness and clarity |
| Reviews and social proof | Partly, if structured | Aggregate rating as one data point |
| Price | Yes, exactly | Price, to the penny |
| Delivery, returns, warranty | Yes, if machine-readable | Certainty of outcome |
Read the right-hand column and a pattern emerges: agents reward operational excellence over marketing excellence. Accurate stock, honest delivery dates, generous and clearly stated returns — the unglamorous things — become competitive assets, because they're the things a machine can actually verify.
Loox
Collects and displays photo reviews building social proof that drives more sales
Best for: Product Reviews
The margin question nobody wants on the agenda
If price-led comparison becomes normal, the obvious risk is commoditisation: your brand equity stops doing the work it used to and you're one row in a table. PwC frames the strategic question for retailers almost exactly that way — what happens to your unit economics if price-first bots become your biggest buyers?
There's a second, more immediate cost. Selling through someone else's agentic checkout is selling through a marketplace, with marketplace economics: OpenAI has been reported to charge merchants a percentage fee on completed in-chat purchases. That may well be worth paying for incremental demand, but it needs to be modelled as a channel with a take rate, not treated as free traffic.
The defences are the ones a machine can see. Exclusive or configured SKUs that don't have an exact comparator. Bundles that change the unit of comparison. Delivery certainty and return terms good enough to win on the fields agents actually weigh. And keeping a direct relationship with the customer, so that a sale mediated by an agent still becomes a repeat sale that isn't — which is why the destination page still matters, and why product page optimisation hasn't stopped mattering just because a machine sometimes reads it first.
→ Is your price and stock accurate in real time? An agent that quotes a stale price and fails at checkout won't come back. Feeds that refresh every half hour break comparison logic.
→ Is your product data complete? Missing attributes don't read as neutral — they read as disqualifying, because the agent can't verify you meet the constraint.
→ Are delivery and returns machine-readable? Buried in a PDF or an image is the same as absent.
→ Can checkout be reached programmatically? Most commerce APIs were built for browser sessions, not machine orchestration.
→ Would your bot defences block a legitimate buying agent? Worth testing before it costs you an order.
Every one of these improves the human experience too. That's the point — it's no-regret work.
The bot paradox
That last check deserves its own moment, because it's a genuinely awkward reversal. Retailers have spent a decade building defences against automated traffic — rate limits, challenge pages, scraper blocking, bot mitigation on checkout. Entirely sensibly: most automated traffic was scrapers, scalpers or fraud.
Some of it is now your customer. Agent traffic looks structurally different from human browsing — bursts of parallel queries against inventory and pricing rather than a leisurely sequential crawl through pages — which is precisely the signature your defences were tuned to catch. The task ahead isn't to open the doors; it's to tell the difference between an agent acting for a real buyer and a bot acting against you, and to make sure the first category can complete a transaction. Very few retailers have that distinction encoded anywhere today.
Agents are not the rational optimisers the theory promises
One important corrective before anyone rebuilds their entire catalogue around price. Academic work evaluating agentic purchasing has found these systems behave less like perfectly rational comparison engines and more like a new species of biased decision-maker, with systematic quirks that differ between models and can be exploited.
Two practical consequences. First, don't assume the cheapest option always wins — presentation of data, ordering, completeness and phrasing all appear to influence outcomes, which is why complete structured data may matter as much as competitive pricing. Second, resist the temptation to game the quirks. Model behaviour changes with every release, and an optimisation built on a specific model's blind spot has roughly the shelf life of a keyword-stuffing tactic in 2011. The same logic we applied to designing websites for AI search holds here: build for legibility, not for exploits.
What's settled, and what isn't
Being honest about the state of play matters, because vendors have an incentive to describe this as further along than it is. What's settled: the protocols exist, the card networks support agent-initiated payment, major retailers are integrated, and referral volume from AI surfaces is growing quickly. What isn't: consumer habit. Most people still prefer to complete a purchase somewhere they already trust, and full delegation of routine buying remains a minority behaviour rather than the norm.
So the sensible posture is neither panic nor dismissal. Do the no-regret work in the audit above, because clean data, accurate stock and clear terms pay for themselves in human conversion regardless of what agents do. Model the commercial questions before someone else's take rate becomes a line in your P&L. And treat this as one more layer of intermediation between you and your customer — the same structural shift we described when creators rather than platforms started driving social sales, and the one marketers are already navigating internally as agentic AI takes over campaign execution.
If you'd rather have the data foundation, the feed quality and the commercial modelling handled as one piece of work, that's what e-commerce marketing support is for.
The practical read
Assistants advise and hand the shopper back to you; agents decide and may never show them your page at all. The rails for that second world are now built, even though the habits haven't caught up, which makes this an unusually comfortable moment to prepare — the work required is the work you should have done anyway. Accurate stock, complete structured attributes, honest delivery dates, plainly stated returns and a checkout a machine can actually reach will all improve human conversion tomorrow and agent selection whenever it arrives. Beyond the plumbing, the strategic question is sharper: if you can no longer win on persuasion, what makes you selectable? For most retailers the honest answer is operational excellence and something that isn't directly comparable — and both of those take longer to build than a feed integration, which is exactly why it's worth starting on them now.
Is your catalogue ready to be chosen by a machine?
We get product data, feeds and checkout ready for agent-led buying.
Explore E-commerce Marketing →Frequently asked questions
What's the difference between an AI shopping assistant and an AI shopping agent?
An assistant advises a person who then buys; an agent acts on that person's behalf within constraints they set, and can complete the purchase itself. The distinction is commercial, not semantic: an assistant still delivers a human to your product page where design and copy can work on them, while an agent may evaluate you entirely through structured data and never load the page.
How do AI shopping agents choose which products to recommend?
Largely on machine-readable attributes — price, availability, delivery timing, specifications, return terms and aggregate ratings — weighed against the constraints the shopper set. Visual design, urgency messaging and brand storytelling carry little weight because an agent doesn't experience them. Research also suggests agents aren't perfectly rational comparison engines and carry systematic biases, so results vary by model.
What do retailers need to do to prepare for agentic commerce?
Prioritise work that pays off either way: accurate real-time pricing and stock, complete structured product data, clearly stated delivery and returns terms, and checkout reachable programmatically. Then review the commercial questions — marketplace transaction fees, and how you defend margin if price-led comparison becomes normal. Finally, check your bot defences aren't blocking legitimate buying agents.
Are AI agents actually buying things yet?
The infrastructure is live and expanding, with open protocols for agent-initiated checkout emerging from late 2025 and card networks adding support for agent-initiated payments. Consumer habit is moving more slowly than the plumbing, and most shoppers still complete purchases on surfaces they already trust. The practical position: the rails exist, so readiness is a reasonable investment even while volumes stay modest.
