Agentic AI in E-commerce: Personalization, Conversational Commerce, and the KVKK Balance (2026)

TL;DR — Agentic AI in e-commerce is no longer a "someday" promise; it is a concrete reality I now meet in almost every project. Shopping and personal-assistant agents, conversational commerce, autonomous merchandising, dynamic pricing, cart-recovery agents, and RAG-powered recommendations over your product catalog are already moving conversion rate, average order value, and retention in visible ways. Gartner predicts that 40% of enterprise applications will embed AI agents by the end of 2026. But if you operate in Turkey, you have to balance all of this against KVKK, Turkey's data protection law; and if you sell to EU customers, the EU AI Act applies on top. In this post I will explain the technical architecture in plain language and show you how to manage real risks — over-personalization "creepiness," discriminatory pricing, hallucinated product claims, and consent fatigue — with concrete safeguards. I will close with a use-case table mapping benefits, KVKK risks, and mitigations, plus the metrics you actually need to track.

I have spent years advising and training companies on artificial intelligence here in Turkey. Over the past year, in nearly every e-commerce team I sit down with, I hear the same sentence: "We need an agent." Sometimes they say it without knowing exactly what they mean; sometimes they point to a chat assistant they saw on a competitor's site. But the underlying instinct is right. The weight of the relationship a brand builds with its customers has grown beyond what human teams alone can carry.

Let me be honest with you. The phrase "agentic AI" has been used so much lately that its meaning has worn thin. So let us clarify from the start. What I mean is not a simple chatbot that answers a question. What I mean is software that, given a goal, plans its own steps to reach that goal, calls tools (catalog queries, stock checks, order systems, payment), evaluates the result, and retries if needed. In other words, not just something that "talks," but something that "gets things done." That is the difference.

Gartner's forecast captures the pace of this shift well: by the end of 2026, 40% of enterprise applications will embed AI agents. I would encourage you to read that number not as a threat or a hype line, but as a timeline. In the period ahead, using agents will stop being a "competitive advantage" and become a "hygiene factor." Not doing something everyone else does simply means falling behind.

But there is a Turkish dimension to this, and I stress it at every opportunity. You cannot copy the slide decks that come out of Silicon Valley wholesale. Because your customer is the Turkish consumer, your price is in Turkish Lira, and your legal framework is KVKK. Any system built without accounting for these three either drives customers away or leaves you exposed in an audit. That balance is the backbone of this post.

I want to explain this without drowning you in jargon, because that is the approach that works best in my trainings. Picture traditional software: you write rules one by one — "if the customer does this, then you do that." Those rules are fixed. When a new situation appears, the software is stumped, because no rule was written for it.

Agentic AI is the opposite. You give it a goal and a set of tools. You say, "Recover this customer's abandoned cart." The agent works out for itself which steps to take: it looks at the customer's past behavior, checks the stock status of the item in the cart, picks a suitable incentive, writes the message, sends it, and monitors the outcome. For each customer these steps may differ slightly. That flexibility is the strength of the agentic approach.

Of course, that strength has a price: unpredictability. With a rule-based system you know exactly what it will do. With an agent-based system you move into a world where it "usually does the right thing." That is why there are two concepts I will keep returning to: guardrails and human-in-the-loop. Putting agentic AI into production without them is like driving a car with no brakes.

Let us get concrete. I will walk through the main use cases I see in the field and recommend to companies, one by one. For each, we will talk about how it works, what benefit it brings, and where you need to be careful.

Shopping and personal-assistant agents

Think of this as a store associate walking beside your customer. In a physical store, a good sales associate asks questions: "Who is it for, what is your budget, what style do you like?" Then they bring you two or three options and explain the differences. A shopping-assistant agent tries to recreate exactly this experience online.

When a customer says, "I am looking for a winter coat under 2,000 TL that handles the cold," the agent searches the catalog, finds suitable products, explains the differences, and can even look at past orders to advise on sizing. Because we have a strongly mobile-first shopping culture in Turkey, talking to an assistant like this instead of scrolling through categories on a small screen can meaningfully speed up conversion.

Conversational commerce

Let us talk about the reality in Turkey: people here shop on WhatsApp. "Is this in stock?", "When will shipping arrive?", "How do returns work?" — a large share of these questions comes through messaging channels. Conversational commerce aims to complete the entire buying journey inside a conversation: finding the product, answering questions, adding to cart, sending the payment link.

What separates this from a traditional FAQ bot is that the agent can actually do things. It does not just say "here is our return policy"; it can find the customer's order and start the return process. I call this not a "talking storefront" but a "talking store." The difference is having real access to the back-end systems.

Autonomous merchandising

Merchandising — the decision of which product goes where and how it is displayed — has classically been the work of experienced teams. Autonomous merchandising agents take over part of these decisions: which product to feature on the homepage, which category to push into a banner, how to highlight items that are running low. The human team sets the strategy; the agent runs the daily micro-optimizations.

Dynamic pricing

I have to be very careful here, because this is both the most powerful and the most risky use case. Dynamic pricing means automatically adjusting price according to demand, stock, competitor prices, and time — like an airline ticket changing over the course of a day. Set up correctly, it improves margin and inventory turnover.

But let me plant a warning immediately: changing price by person — showing the same product to customer A at one price and customer B at another — is a minefield both ethically and legally. I will return to this in detail in the risk section. For now, keep this in mind: product-based dynamic pricing is one thing; person-based discriminatory pricing is something else entirely.

Cart-recovery agents

The abandoned cart is e-commerce's oldest wound. The customer adds the item to the cart and leaves before paying. A cart-recovery agent detects that moment of abandonment and calls the customer back with the right message at the right time. In the classic approach, everyone got the same "your cart is waiting" email. In the agent approach, a hypothesis is formed about why the customer left: was the price too high, did the shipping fee deter them, or were they simply distracted? The message is shaped accordingly.

RAG-powered personalized recommendations

Now we come to the heart of it. Recommendation systems are not new; the "people who bought this also bought that" logic has been around for years. What is new is combining these recommendations with language models and RAG. I will explain RAG in detail in the next section, but briefly: it is the method that lets the agent read your product catalog in real time and speak only from your actual products, without making things up.

In my trainings I say this: if you cannot explain a technology to your boss or your legal team, you are not ready to put it into production. So let me explain these concepts as plainly as I can.

What is RAG (retrieval-augmented generation over the catalog)?

Language models have a well-known weakness: they can confidently make up things they do not actually know, in a tone that sounds certain. We call this "hallucination." In e-commerce this is a disaster; a system that tells a customer about a product that does not exist, a feature that does not exist, or a price that does not exist both loses the sale and exposes you legally.

RAG exists to solve this. It stands for "retrieval-augmented generation." The logic is this: before the model produces an answer, it goes to your own data source (product catalog, stock, price, descriptions), pulls the real information relevant to the question, and grounds its answer solely in that information. In other words, you are telling the model, "Do not talk off the top of your head; look in the cupboard first, and say only what is actually there."

I use this analogy: a language model without RAG is like a very talkative intern who has never seen your products. A model with RAG is like an advisor who has been handed the current catalog and told, "Say only what is written here." In e-commerce, you want the second one.

How recommendation systems work

I explain recommendation systems through three main logics. First, content-based recommendation: it suggests products similar in features to the one the customer is looking at. If they are viewing a pair of red sneakers, other similar sneakers come up. Second, collaborative filtering: the "customers who behave like you also bought these" logic. Third, hybrid models that combine these two and more.

What the agentic approach adds here is that the recommendation stops being a static list and becomes part of a dialogue. When the customer says "that is a bit expensive," the agent can update the recommendation and say, "here is a more affordable alternative." The recommendation is no longer a one-way prediction but a two-way conversation.

Guardrails

A guardrail is a limit you place on what the agent can do. Let me give concrete examples. The agent can never offer a product that is out of stock. The agent can never discount below the floor price you set. The agent can never repeat the customer's identity or card details in text. The agent can never give medical or legal advice. Each of these is a guardrail.

Think of guardrails as the brakes and seatbelt in a car. It is nice for the car to go fast, but a fast car with no brakes just means a crash. In every agent project I put into production, before the "what it can do" list, I have the team write the "what it can never do" list. Because that is where real safety comes from.

Human-in-the-loop

There is pressure for the agent to handle everything on its own, I understand that. But my field experience points the other way: even the most mature systems hand off to a human in certain situations. A high-value return request, an unusual order pattern, a conversation where the customer is angry, a commitment that could carry legal consequences. In these "edge cases," the agent stops and consults a human. We call this human-in-the-loop.

Do not see this as a weakness. On the contrary, it is a sign of maturity. A well-designed system knows where it cannot be certain and, at that point, gracefully passes to a human. For customer experience, too, this often produces a better outcome.

Now we come to perhaps the most critical section of this post. Everything I described above is technically possible. But if you do business in Turkey, "possible" and "legal" are not the same thing. KVKK — the Personal Data Protection Law — is the ground on which all your personalization ambition must rest.

You can think of KVKK as Turkey's parallel to GDPR. The logic is similar, the spirit is the same: personal data belongs to the person; you may process it only on specific, lawful, and transparent grounds. What we call personalization is, by its nature, the processing of personal data. So we are at the very center of KVKK.

Lawful basis and explicit consent for profiling

Personalization, in technical language, means "profiling." You draw a profile from the customer's behavior and make decisions accordingly. Under KVKK, this must have a lawful basis. In most personalization scenarios that basis is explicit consent: the customer knowingly and freely approving that their data will be processed for this purpose.

The key word here is "explicit." A pre-checked box hidden at the bottom of the page is not explicit consent. The customer must clearly understand what they are consenting to, with which data, and toward what outcome. The most common mistake I see in the field is treating consent as a formality. Yet consent is the legal foundation of your personalization; if it is rotten, everything you build on it collapses.

Transparency in automated decision-making

KVKK pays special attention to decisions with significant effects on a person being made by fully automated systems. In other words, if an agent makes a serious decision about a customer without human involvement, the customer has the right to be informed of it and to object. Transparency here is not a technical preference but a legal obligation.

In practice this means: you must be able to answer the question "why am I seeing this recommendation" or "why did I get this price." A black-box system does not just unsettle the customer; it leaves you defenseless before KVKK.

Data minimization and processing limits

This is KVKK's most practical and most overlooked principle: process only the data you need, only as much as you need. Teams usually think "the more data we collect, the better we personalize." This is both legally wrong and dangerous from a risk standpoint. Every extra piece of data you collect is a liability in the event of a breach.

I recommend an approach that flips this around: for each data field, ask "why am I collecting this, and which personalization output does it serve?" If your answer is not clear, do not collect that data. The leaner you get, the easier compliance becomes and the faster the system runs.

If you sell to the EU: the EU AI Act

There is also this: if you sell to customers in the EU, the EU AI Act comes into play. Its logic is market-access based; that is, wherever your company is located, if you offer products or services to the EU market, the rules apply. For many e-commerce brands selling from Turkey into the EU, this is a concrete reality. Depending on the risk class of your AI systems, you may face additional transparency and governance obligations. I will not go into detail, but I have to tell you it should be on your radar.

I always tell the teams I advise: if you are not talking about a technology's risks, you do not know it well enough. Let us discuss the four big risks of agentic AI in e-commerce and concrete measures for each.

1. Over-personalization "creepiness"

There is a line. On one side of it, the customer is delighted: "wow, they recommended exactly what I wanted." On the other side, they shudder: "are these people watching me?" That second feeling is what we call "creepy" — the discomfort created by over-personalization.

The classic example: inferring a life event (pregnancy, a breakup, an illness) before the customer has even acknowledged it themselves and showing ads accordingly. Technically impressive, humanly unsettling. The measure is this: base personalization on information the customer has explicitly given you, not on sensitive conclusions you have inferred. And always give the customer control: "why am I seeing this recommendation" and "turn this off" options should be everywhere. A sense of control largely dissolves the discomfort.

2. The fairness and legality of person-based discriminatory pricing

I touched on this in the dynamic pricing section; let me open it up. Offering the same product to different people at different prices by inferring the customer's ability to pay or their desperation is very dangerous, both in terms of fairness and law. For example, seeing that a customer connected from an expensive phone and showing them a higher price invites a discrimination argument and can conflict with KVKK's principle of fair processing.

The measure: derive price from the product, demand, stock, and time — not from a person's protected characteristics. Keep your pricing logic auditable. Your answer to "why did this customer see this price" must be an answer you can comfortably explain to a judge or an auditor. If you cannot explain it comfortably, do not build that logic.

3. Hallucinated product claims

I touched on this when explaining RAG. An agent telling a customer about a nonexistent feature, a fake warranty, or wrong ingredient information both destroys consumer trust and can create legal liability as a misleading commercial statement. An agent that says "this cream will completely remove the spots on your skin in two weeks" can bring you a serious lawsuit.

The measure is three-layered: first, use RAG to ground the agent solely in real catalog data. Second, in sensitive categories like health and safety, use guardrails to forbid the agent from generating claims entirely; let it only relay the manufacturer's official text. Third, hand off risky conversations to a human via human-in-the-loop.

4. Consent fatigue

This is a more insidious risk. If you show the customer so many "do you accept?" pop-ups, people start hitting "accept" without reading. We call this consent fatigue. The result is ironic: the approvals you added to look compliant actually render consent meaningless. In an audit, your defense "but the customer consented" rests on rotten ground if the customer consented without reading.

The measure: reduce the number of consent requests and make each one genuinely meaningful and contextual. At the moment the customer truly wants to do something, ask only for the permission needed for that moment. Reducing the number of approvals paradoxically strengthens compliance, because the remaining approvals carry real intent.

I use the table below as a starting map in the workshops I run with teams. I recommend adapting it to your own situation.

Use Case	Benefit	KVKK / Legal Risk	Mitigation
Shopping-assistant agent	Higher conversion, better product matching	Consent needed for behavioral profiling	Explicit, contextual consent; "why this recommendation" transparency
Conversational commerce (WhatsApp etc.)	Fast sales, lower support load	Processing of messaging data, channel permission	Channel-based explicit consent; keep only data needed for the transaction
Autonomous merchandising	Better storefront, inventory turnover	Usually low; little personal data	Use aggregate behavior data instead of personal data
Dynamic pricing (product-based)	Margin and inventory optimization	Risk of sliding into person-based discriminatory pricing	Derive price from product/demand, not from the person; keep auditable
Cart-recovery agent	Recovered revenue, higher conversion	Unpermitted contact, aggressive follow-up risk	Contact permission; frequency cap; easy opt-out
RAG-based personal recommendations	Highly relevant suggestions, higher AOV	Profiling + hallucination risk	Real catalog data; data minimization; guardrails
Personal assistant (history-based)	Loyalty, repeat purchase	Sensitive inference / "creepiness"	Rely only on explicit data; give the user a control panel

Once you fill in this table, you will see much more clearly how much risk each scenario carries and where to invest first. My recommendation is usually this: start with low-risk, high-return scenarios (autonomous merchandising, product-based recommendations), and as you settle trust and infrastructure, move to more sensitive ones.

After putting a system into production, saying "it seems to be going well" is not enough. You need to speak with numbers. There are four core metrics that show whether your agentic AI investment is actually working; let me explain each plainly.

Conversion rate. The share of people arriving at your site or chat channel who actually buy. If your agent works, this number should go up. But beware: if you raise conversion while lowering customer satisfaction, the short-term gain is a long-term loss.

Average order value (AOV). How much the customer spends on average per order. A good recommendation system and smart cross-selling raise this number. If the customer buys one more genuinely useful item alongside what they came for, both they are happy and AOV rises.

Cart-recovery rate. The share of abandoned carts you win back. This is the direct report card of your cart-recovery agent. The thing to watch here is not confusing recovery with aggression; short-term recovery achieved by overwhelming the customer pushes you into the "annoying brand" position.

Retention / loyalty. The rate at which a customer who bought once comes back. In my view this is the most valuable metric, because it is the real source of sustainable growth. Well-built, trust-based personalization brings people back. Personalization that unsettles and erodes trust makes the first sale but loses the second.

I recommend reading these four metrics together. If you only look at conversion, you can look good in the short term by overwhelming customers while collapsing retention in the long term. Watching the metrics as a whole protects you from this trap.

Let me turn everything we have discussed into a concrete action plan. If you are an e-commerce brand and you are saying "all right, I am convinced, where do I start," here is the order I recommend to teams.

First, fix your data and consent infrastructure. Before building agents, you must clearly know which data you hold under which consent. This feels like a tedious chore, but it is the foundation you will build on. If your consent records are messy, even the brightest agent cannot protect you in an audit.

Then start with a low-risk pilot. Do not hand your whole store to an agent at once. Set up a pilot in a single category, with a limited audience and clear guardrails. Measure, learn, see the mistakes. Every successful project I have seen in the field started small and grew; projects that started big and shrank usually ended in loss of trust.

Put guardrails and human-in-the-loop in from day one; do not leave them for later. The "make it work first, add safety later" approach is very costly in e-commerce. Brakes are fitted before you speed the car up.

Use transparency as a marketing opportunity. The Turkish consumer is increasingly conscious about data. A brand that says "here is how we protect your data, here is how we personalize recommendations, and you can turn it off anytime" stands apart from competitors who unsettle people. See KVKK compliance not as a burden but as a tool for building trust.

Finally, train your team. A team that builds an agent but does not understand how it works will not know what to do in the first crisis. People understanding these systems, trusting them, and being able to intervene when needed matters more than the technology itself. I have been doing this work for years, and I have seen again and again that the thing that makes the most difference is not the tool but the maturity of the team using it.

If I gather all of this into one sentence: agentic AI is the most powerful tool we have for deepening the relationship you build with customers in e-commerce; but you win only if you use that power within the limits drawn by KVKK in Turkey — and by the EU AI Act if you sell to the EU — keeping the human at the center, measuring, and operating with transparency. Otherwise, even the brightest technology will bring you lawsuits instead of trust, and discomfort instead of loyalty. The choice is hidden in the decisions you make while building the system.