# Generative AI and Agentic Commerce in E-Commerce (2026)

> Source: https://sukruyusufkaya.com/en/blog/e-ticarette-uretken-yapay-zeka-agentic-commerce-2026
> Updated: 2026-07-01T15:53:29.731Z
> Type: blog
> Category: yapay-zeka
**TLDR:** Personalization now drives 45% of online conversions; conversational shopping agents and agentic commerce are rewriting retail. What it means for Turkish e-commerce.

**TL;DR —** E-commerce quietly changed hands in 2026: the customer no longer types keywords into a search box and gets lost among hundreds of products; they shop by talking to an AI assistant. From what I see in the field, most Turkish retailers still think at the level of "let's build a recommendation engine," while globally the game has shifted to "generative commerce" and "agentic commerce" — AI agents that carry out shopping on your behalf. Today roughly **45%** of online conversions come from AI-driven personalization; agentic commerce is expected to contribute **more than $190 billion** to e-commerce revenue by 2030, with roughly **55%** of online purchases potentially made partially or fully through autonomous buying features. In this article I walk through use cases like conversational shopping, hyper-personalization, virtual try-on, dynamic pricing, and returns/support automation — and risks like KVKK compliance, hallucination, brand safety, and agent checkout security — from inside the reality of the Turkish market (Trendyol/Hepsiburada scale, TL pricing, marketplace dynamics).

## The search box is dying: from "type-and-scroll" to conversational shopping

For years the main screen of e-commerce was always the same: a search box at the top, an infinite product list below. The customer types "black sneakers," gets four hundred results, wrestles with filters, gets lost across pages, and often leaves without putting anything in the cart. We have all lived this experience; I remember abandoning exactly such a purchase from a laptop in a hotel room the night before a training session.

In 2026 this picture has changed. Instead of typing keywords into a search box, the customer now talks to an AI assistant: "I'm going to a wedding next week, my budget is 3,000 TL, can you suggest something dark but not too formal?" The assistant understands the intent, context, and budget in that sentence; it lets you explore products conversationally and even visualizes the product on different body sizes and types through virtual try-on. This is called "generative commerce," and it is replacing classic "type-and-scroll" shopping.

This is not a cosmetic change. The search box forced the customer to "speak the machine's language": finding the right keyword and selecting the right filter was the customer's job. In generative commerce, the machine speaks the customer's language instead. The burden shifts from the customer to the platform. And the numerical counterpart of this difference is very clear:

> In 2026, roughly **45%** of online conversions are attributable to AI-driven personalization. That means in nearly one out of every two sales, there is an AI layer in the background understanding the customer and catching the right moment.

Whenever I discuss this shift for Turkey specifically, I always remind people of one point: our market is marketplace-heavy. At the scale of platforms like Trendyol and Hepsiburada, millions of products, tens of thousands of sellers, and an enormous pool of behavioral data accumulate. This scale is both a big opportunity and a big responsibility for generative commerce. An opportunity, because the data to feed the models is already there. A responsibility, because that data is personal data under KVKK and you cannot use it however you like.

## Agentic commerce: AI is no longer an "advisor" but a "buyer"

Now let's come to the most confusing but most critical part. We understood personalization and conversational shopping; but what does "agentic commerce" mean? The difference is this: until now, AI was an advisor that recommended products to you. In agentic commerce, AI becomes an active partner — sometimes even a direct buyer — that carries out shopping on your behalf.

Let me make it concrete. In the classic scenario, you notice your coffee capsules are running out, go to the site, find the product, add it to the cart, and pay. In the agentic scenario, you say something like: "My coffee capsules are almost out, order the usual brand, but this time buy two boxes if it's on sale." The agent tracks your stock, watches the price, catches the right moment, creates the order, and — within the limits you have set — completes the payment. You are not clicking through each step; you state the intent, and the agent handles the rest.

The scale of this scenario is much larger than you might guess:

- Agentic commerce is expected to contribute **more than $190 billion to e-commerce revenue by 2030**.
- In the same period, it is projected that roughly **55%** of online purchases may be made partially or fully through autonomous buying features.
- We saw the first concrete signal of this trend in the past global holiday shopping season: AI and agents influenced roughly **$262 billion** in spending through personalized recommendations and conversational service — which corresponds to roughly **20%** of total holiday spend.

When I share these numbers in a training session, most managers' first reaction is "this is far away from us." And I say: it's not far, it's just invisible. When your customer today tells an AI assistant "find me the top three products in category X," does your product make it into that list of three or not — that is where the real competition now lies. After search engine optimization (SEO), next in line is "agent optimization": your product data needs to be correctly understandable, comparable, and recommendable by AI agents.

## Real-time signals: AI reads the customer "in the moment"

What separates hyper-personalization from old-style segmentation is that it is not static. We used to divide customers into rough groups: "ages 25-34, female, Istanbul, middle income." This segment was defined once and stayed that way for months. AI-driven personalization does not work like that. The system continuously interprets real-time signals and adapts to the customer in the moment.

Which signals am I talking about? Here is the list I go through in the field:

- **Browsing behavior:** Which products they looked at, how long they stayed, what they skipped, what they opened again and again.
- **Purchase history:** What they bought before, how frequently, in which price band.
- **Stated and inferred intent:** A search for "hiking shoes" is stated intent; but browsing the outdoor category for a long time at 11 PM may be an inferred signal of an upcoming holiday.
- **Timing:** Time of day, day of the week, payday period, campaign calendar.
- **Device context:** Mobile or desktop, app or web — even connection speed changes the experience.

The critical point is this: AI combines these signals in real time and adapts to the customer in the moment. The same customer sees one experience while quickly glancing from their phone on the morning metro, and a different experience while researching in detail from a desktop at home in the evening. That is exactly the "hyper" part of hyper-personalization: not a fixed profile, but a moment-to-moment adaptation.

But here is a warning I underline specifically for every retailer in Turkey. Most of these signals are personal data under KVKK. Processing browsing behavior, purchase history, and inferred intent to build a profile relates directly to KVKK's dimensions of "profiling" and "explicit consent." So you cannot just say "we collected the data, let the model use it"; you must clarify which data you process, for what purpose, and on what legal basis. I will return to this topic under a separate heading shortly.

## Use cases in the field: where to start, what they're good for

Let's set theory aside and move to concrete use cases, because this is the question I get most in trainings: "Fine, but where do we start?" The table below summarizes the concrete use cases of generative and agent-based AI in e-commerce, what they do, and their counterparts in the Turkish context.

| Use case | What it does | Turkey/field counterpart |
|---|---|---|
| Product discovery and search | Intent-based, conversational search instead of keywords | Solves finding the right product among millions on a marketplace |
| Conversational assistant | Dialogues with the customer, asks questions, guides | 24/7 Turkish shopping advisor; reduces call center load |
| Personalized recommendations | A "just for you" storefront driven by real-time signals | The main lever feeding ~45% of conversions |
| Dynamic pricing | Price adjustment by demand, stock, competition, and timing | Powerful but risky under TL volatility and heavy campaigns |
| Generated product descriptions and ad creative | Automated text and visuals for thousands of products | Solves the scale problem on high-seller-count marketplaces |
| Review summarization | Distills hundreds of reviews into a single summary | Fast answer to "what do people mostly complain about?" |
| Returns and support automation | Runs returns, exchanges, and support requests end-to-end | Lowers operational cost in high-return categories |
| Inventory and demand forecasting | Predicts the future from sales signals | Balances stock-outs and overstock in campaign periods |
| Virtual try-on | Visualizes the product on different sizes and body types | One of the strongest tools for lowering returns in fashion/cosmetics |

Whenever I share this table, I always add: don't try to do all of it at once. In the field, the most successful organizations were the ones that started with a single narrow problem. For example, if the return rate is high in the fashion category, it is much smarter to first focus on virtual try-on and review summarization — to lower returns and clearly demonstrate return on investment. A powerful yet risky area like dynamic pricing should be entered last, in a controlled way.

## Conversational assistants and generated content: the price of scale is hallucination

Conversational assistants and generated content (product descriptions, ad copy, visuals) are the most visible AI uses in e-commerce. Writing descriptions by hand for the products of tens of thousands of sellers on a marketplace is impossible; generative AI can solve this scale single-handedly. And an assistant that can talk to the customer in Turkish, naturally and in context, directly affects conversion.

But this is exactly where the most painful risk in the field kicks in: **hallucination.** A generative model can write a nonexistent product feature as if it were real. It says "this shoe is waterproof" when it isn't. It says "this phone has a 5-year warranty" when it's 2. The customer trusts this information and buys; when the product arrives, they face reality, and the result: returns, complaints, reputational damage, and — in Turkey — serious liability under consumer law.

I once summarized this in a training session like so:

> Generative AI is like an intern writing product descriptions: it's fast, it doesn't tire, it finishes thousands of products the same day. But you would never publish what it writes without an editor verifying every sentence. In e-commerce, that editor must be a verification layer bound to the source (the product database).

The practical precautions are these:

- **Grounding to the source:** The model should generate product features not from its own "memory" but from a verified product database. Not free generation, but data-driven generation.
- **Brand safety filters:** Generated text and visuals should be automatically checked for compliance with the brand's tone of voice, legal limits, and advertising regulations. In Turkey, matters like health claims and price promises are additionally sensitive.
- **Human approval layer:** Automation in high-volume but low-risk categories; mandatory human approval in sensitive categories like health, baby products, and food.

I emphasize brand safety in particular because generated visuals and text can sometimes misrepresent the brand unintentionally or produce competitor/inappropriate contexts. Imagine a marketplace: tens of thousands of product descriptions are generated automatically. If just one of them contains a false health claim or an exaggerated promise, it can put the whole brand at risk. That's why generated content must be designed with automation for speed and with oversight for trust, together.

## Dynamic pricing: powerful under TL volatility, but the most sensitive tool

I want to address dynamic pricing under a separate heading, because in the Turkish context this is both the area with the most potential and the one demanding the most care. By reading demand, stock levels, competitor prices, time of day, and customer signals, AI can adjust the price in the moment. Given TL volatility and frequent campaign cycles, this capability is extremely valuable for protecting margins.

But this very power brings the biggest risk: **fairness in pricing.** What happens if AI shows the same product to two different customers at different prices? If this difference is based on inferences about the customer's ability to pay, it creates a serious problem, both ethically and reputationally. When customers notice this kind of differentiation, the reaction is sharp; a single screenshot on social media can tear down the trust you built over months.

The approach I recommend in the field is clear:

- Base dynamic pricing on legitimate, explainable signals like **supply-demand, stock, and competition**.
- Avoid personalizing the price based on **inferences about an individual customer's ability to pay**; this conflicts both with KVKK's profiling dimension and with the basic principle of fairness.
- Keep price changes **explainable**: you should be able to give a sensible internal answer to "why this price?"
- Preserve **transparency** in discounts and campaigns; in Turkey, both regulation and public sensitivity around fake discounts and price manipulation are high.

AI-driven pricing, when set up correctly, can also delight the customer: the right discount at the right moment, clearing excess stock, balancing demand — all to everyone's benefit. Set up wrong, it destroys trust. The difference is hidden in which signals you rely on.

## Returns, support, and forecasting: the invisible heroes of operations

We've talked about the flashy use cases on the storefront; but the real profitability of e-commerce is often hidden in operations, that is, behind the scenes. Generative and agent-based AI has three powerful uses here: returns/support automation, review summarization, and inventory/demand forecasting.

**Returns and support automation** significantly lowers operational cost, especially in high-return categories (fashion above all). An agent can run the customer's return request from start to finish: it understands the reason, checks eligibility, generates the shipping label, and records the return. The customer doesn't wait for hours in a call center; the process finishes in minutes. I describe this in trainings as "invisible satisfaction": when a customer has a good return experience, they rarely praise it, but when they have a bad one, they definitely leave.

**Review summarization** is a tool that looks small but touches conversion directly. Reading hundreds of reviews is tiring for the customer; AI distills them into a single summary: "Users praise the quality but often note that the sizing runs small." That single sentence, by helping the customer choose the right size, both increases conversion and lowers returns.

**Inventory and demand forecasting** predicts the future from sales signals, striking the most critical balance during campaign periods: neither running out of stock and missing sales, nor locking up cash with overstock. Given the intensity of major discount periods in Turkey (November campaigns, end-of-season sales), accurate demand forecasting flows directly into cash flow.

The common feature of these three areas is that they don't catch the customer's eye but directly affect profitability. The talking assistant on the storefront is visible and impressive; but the economics of the business are often determined by these invisible heroes.

## KVKK, trust, and agent checkout security: what must be solved before technology

Now we've reached the section I dwell on the most. Everything I described above collapses without a solid foundation of trust and compliance. When discussing AI in Turkish e-commerce, there are three issues that must be solved before the technology.

**First, the balance between KVKK and personalization.** Hyper-personalization is by nature a data-hungry approach: the more signals, the more accurate the recommendation. But browsing behavior, purchase history, and inferred intent are personal data under KVKK. Processing these to build a profile is a "profiling" activity and is subject to principles like explicit consent, the duty to inform, and purpose limitation. The most common mistake I see in the field is the team thinking "the data is already in our hands, let's feed the model." But having the data in your hands does not mean you can use it for any purpose. If you start without clarifying which data you process for personalization and on what legal basis, one day an audit or complaint could halt the entire project.

**Second, agent checkout security.** The most exciting promise of agentic commerce — the agent buying on your behalf — is also the biggest security question. If an agent pays on your behalf: What's the spending limit? Which transactions require separate approval? What happens if the agent is compromised? How is payment information protected? Enabling autonomous buying without answering these questions is like leaving the door open and saying "I hope no one comes in." I recommend to organizations: agent payments should have clear spending caps, human-approval thresholds by transaction type, and a traceable record of every agent transaction. Autonomy does not mean limitlessness; it means freedom within well-defined limits.

**Third, brand trust.** The AI-driven experience is built on the customer's trust in the brand. Giving wrong product information through hallucination, doing unfair pricing, or misusing data — each of these erodes trust. And in e-commerce, trust is the most expensive but most fragile asset. Once lost, regaining it is far harder than earning it.

The practical way to think about these three issues together is in this table:

| Risk area | Concrete danger | Precaution to take |
|---|---|---|
| KVKK / personalization | Unauthorized profiling, purpose creep | Explicit consent, disclosure, purpose-limited data processing |
| Hallucination | False product claims, consumer harm | Source-grounded generation, verification layer |
| Brand safety | Inappropriate generated content, false claims | Automated filters + human approval in sensitive categories |
| Agent checkout security | Unauthorized/compromised purchases | Spending caps, approval thresholds, traceable records |
| Fairness in pricing | Discriminatory pricing by ability to pay | Reliance on legitimate signals, transparency |

## Where to start: a practical 90-day roadmap

When I walk organizations through this whole picture, the question is always the same: "So what do we do tomorrow morning?" Let me share a starter plan that works, drawn from the field. The goal is not to do everything at once; it's to build trust with a fast, measurable win and keep going.

1. **First 30 days — pick a narrow problem and clean up the data.** Start with a single high-return category (fashion is usually the best candidate). Clean your product data; because both generative AI and agent recommendation systems work badly with bad data. In the same period, clarify the KVKK side: which data will you process for what purpose, and is your explicit-consent flow ready?
2. **30-60 days — deploy a low-risk, high-impact use case.** Review summarization and source-grounded product description generation are the ideal start: they deliver fast results, carry low risk, and are easy to verify. Virtual try-on shows a clear return by lowering returns in fashion/cosmetics.
3. **60-90 days — pilot the conversational assistant.** Test a Turkish, context-aware shopping assistant in a limited category. Build in hallucination safeguards (grounding, verification) from the start; bind the assistant to the product database, don't let it run free.
4. **90 days and beyond — move into dynamic pricing and agentic features under control.** Enter these powerful but sensitive areas only when your data, compliance, and trust foundation is solid. Do not enable autonomous buying in agent payments without spending caps, approval thresholds, and traceability.

The spirit of this roadmap is this: AI in e-commerce is not a "big bang" project, but a chain of small, measurable wins stacked on top of each other. The transition from the search box to conversational shopping, from advisor AI to buyer agent, has already begun. For the Turkish retailer, the real question is not "will we use AI?"; it's "when our customer tells an agent 'find me the best,' will we be the one making the list?" That first narrow step you take today will determine whether you appear on that list.