The 30-second summary: the AI agents market is moving from answering to acting. McKinsey estimates agentic commerce could orchestrate $3–5 trillion of global retail by 2030, and Adobe measured a 693% jump in retail visits referred by AI tools over the 2025 holiday season. Amid the race between OpenAI, Google, Visa and Mastercard, Technology KSA has shipped a different model through Mosaiden: a marketplace open to every AI assistant, checkout that never leaves the marketplace, and human approval before every action.
What are AI shopping agents?
AI shopping agents are the next generation of AI agents: software that does not stop at answering "what's the best phone under SAR 2,000?" but carries out the task itself. It understands your request, searches store catalogues, compares prices and specifications, then prepares the cart or booking for you to review and pay for yourself.
The difference between traditional shopping assistants and shopping agents is fundamental:
- A shopping assistant suggests: it gives you links and recommendations, and you finish the job.
- A shopping agent acts: it calls real tools (search, price comparison, cart building, booking) and comes back with a result ready for your approval.
That shift from talking to doing is what the industry now calls agentic commerce.
How big is the AI agents market in commerce?
| Indicator | Figure | Source |
|---|---|---|
| Global agentic commerce by 2030 (estimate) | $3 – 5 trillion | McKinsey, Oct 2025 |
| US retail share (estimate) | Up to $1 trillion | McKinsey, Oct 2025 |
| Growth in AI-referred visits to US retail sites (Nov–Dec 2025) | +693% year on year | Adobe, Jan 2026 |
| Conversion advantage of AI-referred visitors over other sources | +31% | Adobe, Jan 2026 |
| Active e-commerce registrations in Saudi Arabia (Q4 2025) | 43,854 (+9% YoY) | Ministry of Commerce via Mubasher |
Read these numbers with care: McKinsey's figures are scenarios that depend on adoption speed, and Adobe's growth rates are high partly because they start from a small base. The direction, though, is clear: shoppers are handing the search-and-compare stage to an AI agent, and a store the agent cannot read will drop out of the decision.
The giants' race: who is writing the rules of AI shopping?
Throughout 2025 the largest technology and payments companies announced their infrastructure for shopping agents:
| Initiative | Company | Core idea |
|---|---|---|
| Agentic Commerce Protocol and Instant Checkout | OpenAI with Stripe | Buy directly inside ChatGPT with a shared payment token scoped to the merchant and basket total |
| Agent Payments Protocol (AP2) | Google and partners | Cryptographically signed "mandates" proving what the user asked for and its limits, built on MCP and A2A |
| Visa Intelligent Commerce | Visa | Tokenised, agent-ready cards with user-set spending limits |
| Mastercard Agent Pay | Mastercard | "Agentic tokens" that make every agent-initiated transaction identifiable and verified |
What they share: most of these models were designed for the US market first, in English first, and many move the moment of payment into the conversation itself. That is exactly the gap a Saudi company picked up.
The gap nobody in the region had filled
A Saudi shopper who asks their AI assistant for "the cheapest iPhone price in Riyadh today" runs into three problems:
- No live Arabic catalogue: global models cannot see current Saudi store prices in riyals, so they answer from stale memory or invent a number.
- Agents in walled gardens: every platform builds its own agent that only sees its own stores, leaving the user locked into one platform.
- Payment trust: typing card details into a chat with a language model is not something users are comfortable with, nor what regulators prefer.
How Technology KSA rewrote the rules: five new rules from Mosaiden
Technology KSA (شركة تكنولوجيا سعودية) built Mosaiden (سوق المساعدين): a Saudi marketplace that brings stores, classified ads, restaurants with their menus, hotels, flights, cars, real estate and services into one place, where AI assistants search on your behalf, compare prices, order and book with delivery. What makes it globally notable is not the size of the catalogue but the rules it was built on:
Rule 1: the marketplace inside every AI, not another closed agent
Instead of asking users to install "yet another agent", the platform opened a public MCP endpoint at mosaiden.com/mcp. The Model Context Protocol (MCP) is an open standard that AI assistants use to talk to external tools, so any assistant that supports it, such as Claude and ChatGPT, can search the Saudi marketplace, compare phone and electronics prices, and browse restaurant menus and hotels in Saudi riyals. The user stays in their favourite assistant; the marketplace comes to them.
Rule 2: payment never enters the chat
When the user decides to buy, the agent does not ask for card details. It builds a cart link from the products in the results; the user opens it, reviews, edits and pays on mosaiden.com itself. The chat is for searching and comparing; payment belongs on the official site. This rule alone removes the biggest trust barrier in agentic commerce.
Rule 3: nothing happens without your approval — preview, then confirm
Every write action from inside the chat, from creating a store to adding a product or publishing a listing, goes through two steps: a preview shown to the user exactly as it will appear, then a confirmation that only happens with their explicit consent. The agent proposes and executes, but the decision always stays human.
Rule 4: agent negotiates with agent
In classified ads, the seller's assistant contacts the assistants of interested buyers and exchanges offers within limits its owner sets, then notifies the owner of the outcome to accept or decline. It is an early picture of where markets are heading: agents negotiate, humans decide.
Rule 5: Arabic first, Saudi content, and merchant text treated as data, not commands
The interface and the assistant work in Arabic, prices are in riyals, and results come from real stores, restaurants and hotels. Technically, the most important part: the platform treats any text written by a merchant or advertiser as data, not instructions, so a booby-trapped product description cannot steer the agent into doing something the user never asked for. That is a direct defence against prompt injection, one of the most serious weaknesses of AI agents.
Comparison: the prevailing global model vs the Mosaiden model
| Criterion | Prevailing global model | Mosaiden model |
|---|---|---|
| Where does the agent live? | Usually inside one platform | Inside any MCP-compatible assistant |
| Where does payment happen? | Inside the chat in many cases | On the marketplace only, with no card details in chat |
| Executing actions | Varies | Preview, then explicit confirmation for every action |
| Language and market | English and the US first | Arabic and Saudi Arabia first, prices in riyals |
| Merchant text | Read as context | Treated as data, never executed as instructions |
| Classifieds fees | — | Free to list; 1% commission only when a deal closes |
How to shop smart with an AI agent: a 5-step guide
- Describe what you need precisely: budget, city, and the specifications you won't compromise on.
- Ask for a comparison explicitly: "compare this product's price across several Saudi stores in riyals".
- Ask for the source: a good agent gives you the original product link for every price.
- Review the cart before paying: open the cart link and check quantity, price and delivery.
- Pay on the official site only: never share card details inside any chat.
What this means for Saudi merchants and businesses
If an AI agent is doing the searching for the customer, a store the agent cannot read will never be shortlisted. Getting ready for the next generation of shopping agents means:
- A structured catalogue with current prices and stock and structured data (Schema) that both search engines and agents understand.
- An MCP endpoint or API that lets AI assistants search and compare without brittle page scraping.
- Optimising for answer engines (AEO) and generative engines (GEO), not only for traditional search.
- Clear trust controls: payment on your own site, explicit approval, and text that is never executed as commands.
Technology KSA brings this experience to businesses that want to be ready when the AI agent becomes their first customer. See also the Mosaiden platform case study, KSA Search and optimising for AEO and GEO, and the role of MCP in intelligent email systems.
The challenges: what could slow shopping agents down?
- Trust and liability: who is responsible if the agent buys the wrong product? That is why human approval before payment is a requirement, not a luxury.
- Price accuracy: an agent answering from memory may "invent" a price. The fix is to connect it to a live catalogue and forbid it from quoting anything not in the results.
- Privacy: data collection must comply with Saudi Arabia's Personal Data Protection Law (PDPL).
- Prompt injection: malicious text in a product description may try to steer the agent; protection starts with treating every external text as data.
Frequently asked questions about AI shopping agents
What are AI shopping agents?
AI shopping agents are AI programs that carry out a purchase rather than just answer questions about it: they understand your request, search stores, compare prices and specifications, then prepare the cart or booking for you to review and pay for yourself.
What is the difference between a shopping assistant and an AI shopping agent?
A shopping assistant suggests and answers; a shopping agent acts. It calls real tools such as catalogue search, price comparison and cart building. A good agent stops before any final action and asks for your approval.
How big is the market for AI agents in commerce?
McKinsey (October 2025) estimates that agentic commerce could orchestrate $3 trillion to $5 trillion of global retail sales by 2030, including about $1 trillion in the US. These are forward-looking estimates, not realised figures.
How is Technology KSA changing the rules for AI shopping agents?
Through Mosaiden (mosaiden.com): instead of building a closed agent, it opened the Saudi marketplace to any MCP-compatible AI assistant, keeps checkout on the marketplace without asking for card details in chat, requires a two-step preview-and-confirm before any purchase or listing, and lets assistants negotiate for buyers and sellers within limits the user sets.
Is shopping through an AI agent safe?
It is safe when the agent follows three rules: it never asks for card details inside the chat, it never buys or publishes without explicit approval, and it treats merchant text as data rather than instructions. Always check that the payment link is on the store's official site.
How do I shop smart with an AI agent in Saudi Arabia?
Describe what you need precisely (budget, city, specifications), ask the assistant to compare prices across several stores in Saudi riyals, review the cart link before paying, and complete payment only on the store's official website.
Sources
- McKinsey — The agentic commerce opportunity (October 2025)
- Adobe — 2025 holiday shopping season results (January 2026)
- Stripe — Instant Checkout in ChatGPT and the Agentic Commerce Protocol
- Visa — Visa Intelligent Commerce
- Mastercard — Agent Pay
- Mubasher — Saudi e-commerce registrations, Q4 2025
Disclosure: Mosaiden is developed by Technology KSA, the publisher of this blog. Market figures are attributed to the sources above, and the platform description is limited to features that are live.
