Agentic Commerce: How AI Agents Buy and Sell

Agentic commerce is the term for what happens when an AI agent, not a person, does the browsing, comparing and buying on a store's behalf. It sounds like a future-tense concept until you look at what happened this past holiday season. Retailers saw a real jump in orders coming from AI referrals, and those orders converted better than normal traffic. This is for store owners, operators and marketers who want to understand what's actually happening, not the hype version, and what to fix before it costs them sales they never see coming.

If you sell through WhatsApp and want the mechanics of getting your catalog into ChatGPT and Claude specifically, we've written about that separately.

What agentic commerce means

It describes AI systems that can research, decide and transact with limited or no human input at each step. A chatbot that answers "do you have this in blue" is not agentic. A system that can take "find me a good running shoe under 3000 rupees, order it, and have it here by Friday" and complete that sequence, including the payment, is.

Three things separate this from the automation ecommerce teams already run:

• The agent interprets unstructured information (reviews, product descriptions, return policies) instead of following a fixed rule
• It can compare across multiple sellers before deciding, rather than executing a single pre-set action
• It acts with some degree of authorisation to complete the purchase, not only recommend one

Most businesses already have automation - reorder triggers, abandoned cart flows, subscription renewals. None of that is agentic in the sense being discussed now. What's new is an outside AI system, one your store doesn't control, making the discovery and purchase decision using your product data as its only source of truth.

That last part matters more than most explainers give it credit for. When a human lands on a confusing product page they can still figure it out, they scroll, zoom into a photo, read a review that clarifies sizing. An agent doesn't. It reads what's there, and if the data is thin or inconsistent it moves on to a competitor with cleaner listings.

Why this is happening now

Adobe's analysis of the 2025 US holiday shopping season found that traffic to retail sites referred by generative AI tools jumped roughly 693% year over year, and that traffic converted about 31% better than traffic from other sources. Retail saw the sharpest AI-referral growth of any industry Adobe tracked, ahead of travel, financial services and media.

What makes that worth paying attention to isn't the size of it, it's the conversion gap. Shoppers arriving through an AI agent were closer to a decision by the time they clicked through. The agent had already done the comparing and narrowing that used to happen across five open browser tabs. By the time a human or an agent reaches your product page in this flow, a lot of the persuasion work is already finished, somewhere you can't see and can't optimise with a landing page test.

This is the shift that matters for operators: discovery is moving upstream, into a conversation you don't have visibility into, before a customer ever reaches your site.

How AI agents shop

Strip away the protocol jargon and an agent buying on someone's behalf goes through roughly three phases, and each one depends on something your store either has or doesn't.

Discovery, can the agent find you at all. Agents don't browse the way people do, they query. That query might hit a search index, a product feed, or increasingly a direct connection to a store's catalog through a protocol like MCP (Model Context Protocol), which lets an AI assistant pull structured product data directly instead of scraping a page. Fufa's catalog integration works this way, exposing a merchant's WhatsApp product catalog so it's discoverable inside ChatGPT, Claude and WhatsApp itself, not just a browser search bar. If your product data lives only as pretty images and marketing copy on a webpage, the agent has to work harder to understand it, and often just skips it for a competitor whose data is structured and unambiguous.

Evaluation, does the agent trust what it finds. Once your product is in front of it, the agent decides whether it's a fit, using the same signals a careful human would - specs, price, availability, return terms, review sentiment. The difference is it applies them consistently and instantly across dozens of options, where a person checks two or three before getting tired of comparing. The deciding factor is rarely which AI model is doing the shopping. It's whether the merchant's underlying data is clean enough for any system, human or agent, to trust without a phone call to support first.

Transaction, who's authorised to pay. This is the part still being worked out industry-wide. Visa, Stripe and others are building tokenised "agent cards", scoped payment credentials that let an agent spend up to a defined limit under defined conditions. For most small and mid-sized stores right now, the practical version looks less like an agent silently completing checkout and more like an agent narrowing the field to one or two options and handing the actual purchase back to a human, often inside a messaging thread on WhatsApp.

That handoff point, where the agent's recommendation becomes a human's confirmed order, is exactly where a well-configured WhatsApp flow earns its keep. A customer who gets a shortlist from an AI assistant and then clicks through to a business's WhatsApp number to confirm size, ask about delivery and pay is already living in a hybrid version of agentic commerce, even if nobody on either end calls it that.

This is also where the excitement gets ahead of reality. Full agent-to-agent transactions, where a shopper's assistant negotiates directly with a merchant's system and settles payment without a human clicking anything, exist mostly in pilot programmes and enterprise B2B procurement. For most consumer retail, agentic commerce today is better-informed discovery followed by a human-confirmed purchase, not a fully autonomous checkout. That's still a meaningful shift. It just isn't the sci-fi version some vendors are selling.

What breaks when your catalog and checkout aren't ready

Take a hotel listing written for humans deciding on vibe - "breezy", "romantic getaway" - with the actual bed configuration and cancellation terms buried three clicks deep. An AI travel agent evaluating that property against a traveler's stated criteria, two queen beds, free cancellation within 48 hours, has nothing structured to check against. It moves on to a competitor with the terms spelled out in the first line. Nothing about the property was a bad fit. The data just wasn't legible to the thing doing the evaluating.

Ecommerce fails the same way. Vague titles, missing attributes, out-of-stock items still showing as available, all of it relies on human forgiveness that doesn't exist for a machine reading a feed. A product an agent can't confidently evaluate is a product that doesn't get recommended, full stop.

Teams treat the product feed and catalog as something to clean up eventually, and it turns into an urgent fire drill the first time an agent-referred order fails at checkout or an agent simply stops surfacing the store.

Checkout friction is the other failure point. CAPTCHAs, multi-step address forms, JavaScript-heavy payment widgets built assuming a human is patiently clicking through - these block legitimate agent-driven buyers the same way they block bots, because from the server's point of view they often look identical.

Getting the store ready for agent buyers

None of this requires rebuilding your stack. It requires treating product data as infrastructure instead of an afterthought.

1. Structure the catalog properly. Consistent attributes for brand, size, colour, material and price. Don't rely on a customer to infer what a vague title means.
2. Keep availability accurate in near real time. An agent that recommends an out-of-stock item once is less likely to trust your feed next time. There's no forgiveness loop the way there is with a human who calls to ask.
3. Reduce friction at the point of purchase. Audit anywhere your checkout assumes a slow, visual, human interaction, and see what breaks when it isn't one.
4. Make your WhatsApp presence part of the discovery surface, not just support. For a lot of Indian D2C and services businesses, WhatsApp is already where the human confirmation happens after an agent narrows the options. If your catalog isn't synced and clean there, you lose that handoff. WhatsApp for Ecommerce covers the setup.
5. Set explicit rules for what gets automated and what needs a human. COD confirmation, high-value orders and first-time customers are reasonable places to keep a person or a verification step in the loop even as more of the funnel becomes agent-driven.

If you want help getting your WhatsApp catalog structured so it works for both human shoppers and AI agents, book a quick call with our team and we'll go through what's actually blocking discovery versus what's a minor fix.

The limits nobody's talking about

Agentic commerce still has real, unresolved problems, and a business betting its whole strategy on agents replacing human buyers this year is going to be disappointed.

Identity and liability are the biggest. If an agent places an order a customer didn't actually want, who's responsible - the platform, the merchant, or the person who authorised the agent? There's no settled legal answer yet, in India or anywhere else. Prompt injection is a real risk too: malicious content on a page or in a document can trick an agent into buying the wrong thing or transacting with a fraudulent seller, and merchants have limited ways right now to prove to an agent that they're legitimate.

Hand full purchase authority to an agent before verification and trust mechanics are sorted, and you end up fielding disputes a simple human confirmation step would have prevented entirely. That's not a reason to ignore agentic commerce. It's a reason to build the human checkpoint in deliberately rather than removing it because the technology sounds finished. Regulatory clarity, payment network standards and agent identity verification are all still catching up to what the AI platforms can technically do. Build for that gap, not around it.

There's a quieter limit worth naming too: not every product category benefits equally from agent-driven discovery. High-consideration purchases, things people want to touch, try, or discuss with someone before buying, still pull humans back into the loop even after an agent does the initial narrowing. Treating agentic commerce as a blanket strategy across a whole catalog, instead of a channel that fits some products better than others, is a common early mistake.

A store that's ready for agentic commerce in 2026 doesn't look radically different from a well-run store today. The catalog is clean and consistent. Pricing and stock are accurate without a two-day lag. Checkout doesn't assume every visitor is a patient human with nowhere else to be. And somewhere in the flow, usually right before payment, there's still a human checkpoint for anything higher-stakes than a routine reorder.

If you're already on Fufa, the next workflow worth configuring after your WhatsApp inbox is live is exposing that same product catalog through MCP so it's discoverable inside ChatGPT and Claude, not just WhatsApp search. The infrastructure isn't experimental anymore. Clean the catalog first. Everything else builds on that.

Things people ask

What is agentic commerce in simple terms?

AI agents doing the shopping instead of, or alongside, a human. The agent researches products, compares them against a stated need, and either completes the purchase or narrows it to one or two options for a person to confirm. It's different from a chatbot answering questions because the agent is taking action toward a transaction, not only responding.

Is agentic commerce the same as chatbot shopping assistants?

Not quite. A basic shopping chatbot answers questions and maybe recommends products within one site. An agentic system can operate across multiple sources, reason about tradeoffs the way a person would (price versus delivery time versus reviews), and in some setups execute the purchase. The reasoning and cross-source comparison is the defining difference.

Do small and mid-sized stores need to worry about this yet?

It's relevant now, and arguably more urgent for smaller stores, since large retailers already have teams dedicated to feed management and structured data. A smaller D2C brand with a messy catalog and a slow checkout is the one most likely to get skipped entirely by an agent doing a quick comparison, simply because the data isn't there to evaluate.

When should we keep a human in the loop?

High-value orders, first-time customers, and anything involving cash on delivery are reasonable places to keep a human confirmation step, at least for now. Liability and identity verification for agent-initiated purchases aren't settled, so the safer approach is letting agents handle discovery and narrowing while a person confirms anything with real financial exposure.

How do small teams prepare without a dedicated data or engineering team?

Start with the catalog, not the technology. Clean, consistent product titles, accurate stock levels and clear attributes go a long way before you need any tooling. Platforms that already structure your WhatsApp catalog and expose it to AI discovery layers, like Fufa's MCP integration, remove most of the technical lift, so the team's real job is keeping the underlying product data accurate.

What's the minimum setup to see a real result?

An accurate, structured product catalog and a checkout that doesn't block non-human traffic with CAPTCHAs or heavy JavaScript gates. That combination alone decides whether an agent can find, trust and complete a purchase involving your store. Everything else - negotiation protocols, agent-specific payment cards, deeper personalisation - builds on that.

Can AI agents negotiate prices with sellers?

In B2B, yes, this is starting to happen through structured offer and counteroffer protocols, though it's early. In consumer retail most pricing is fixed, so the "negotiation" an agent does is closer to finding the best available price or promotion across sources than haggling with a merchant.

Will agentic commerce replace human customer support and sales conversations?

No, and treating it that way tends to backfire. Automate too much too early and you get a bot that handles nothing particularly well. Agents are good at narrowing options and handling routine, well-defined transactions. Judgement calls, disputes and anything emotionally charged still need a person, and building that boundary in deliberately produces a better experience than automating around it.