Agentic Commerce: How AI Agents Run Your Store

In Brief

Agentic commerce is a way of running a store where AI agents make the daily operating decisions instead of reporting them to you. Inside a defined goal and permission ceiling, the agent acts on its own: it opens a purchase request when stock drops, cuts budget on a campaign whose conversion is sliding, messages the customer whose parcel is late. You set the rules and the limits, then audit the log. Getting started needs no new dashboard, only an API and a permission model an agent can safely use.

It is 22.40 on a Friday. The last seven units of one model sold in two hours, the product still sits first in your storefront, and Google Ads keeps spending 1.200 TL a day on it. When you check on Monday, see zero stock and pause the campaign, you have paid roughly 3.000 TL over two and a half days for something nobody could buy. Your reporting screen knew this on Friday night. Nobody was looking.

The missing piece is not data. It is authority. Most platforms produce the right number at the right moment, then wait for a person to press a button. At midnight, on a holiday, and on the campaign day when orders peak, that person is elsewhere.

Agentic commerce closes the gap by turning the AI from the eye of the system into its hand. It reads the data, checks it against your rules and acts inside the boundary you drew. For the sales side of the same shift, the piece on how agentic e-commerce changes daily operations pairs well with what follows.

What agentic commerce is, and how it differs from rule-based automation

Agentic commerce means running commercial processes around a stated goal rather than a list of instructions. You do not say "if stock falls below five, do X." You say: keep this category's out-of-stock rate under 2 percent at no more than 180 TL landed cost per unit. The agent picks the tools, measures the outcome and retries when it misses.

Where a rule list breaks down

Classic automation thinks one condition at a time. Stock fell below five, the rule fired, the email went out. In reality the supplier may be on holiday, the carrier two days late in that district, the ads still live, and the same stock pool selling on a marketplace. Every combination written as its own rule becomes a list nobody maintains past month three.

An agent reads those conditions as one situation rather than four rules. Seeing the supply delay, it does not kill the campaign. It drops the daily budget from 1.200 TL to 300 TL, locks the marketplace allocation and keeps the product live with a preorder label. Three actions across three systems, one decision.

Three levels of maturity

• Analytical AI: forecasts from history, then leaves the action to you.
• Generative AI: writes the product description and suggests a headline, then waits for approval.
• Agentic AI: applies its own output. It writes the description, saves it to the product, refreshes the sitemap and measures the result a week later.

What separates them is not model quality but connectivity and permission. The same model without cancellation rights is a chat box. Connected to your order API with a ceiling, it is an operations hire.

What an agent actually runs on: goal, tools and permission

An agentic setup has three parts: the goal it optimises, the functions it may call, and the permission matrix separating what it does alone from what it escalates. Miss one and you do not have an agent, you have a script with unusually broad access.

Function calling and MCP

A model needs a standard interface to reach your systems. Model Context Protocol is the contract that lets it call functions such as "update product stock" with typed parameters. The model writes no SQL and never touches the database. Cap your price function at 15 percent and no amount of clever prompting produces a 40 percent discount.

A sample permission matrix

Add one more column when you write your own: whether the action is reversible. A price change rolls back, a sent SMS does not. Irreversible actions belong on the escalation side.

Keep the table narrow on day one, log every action and have one person read the log weekly. Widen limits as trust builds. Starting broad and tightening after an incident costs more than it saved. The permission layer is laid out tool by tool on the Tecof AI agents page.

What it returns on pricing, stock and ad budget

The payback shows up here first, because these three systems are linked and a person usually touches only one. Take a store with 400 orders a month and an 850 TL average basket. The table below is a template, not a benchmark. Swap in your own numbers.

The biggest line is not ad spend. It is the days you could not sell. Six days out of stock is close to a fifth of that product's monthly revenue, and it never appears as a cost anywhere.

Dynamic pricing that protects margin

The common failure of a pricing agent is treating competitor price as the only input. A rule that says "always one lira cheaper" drags you into the hole when a rival dumps stock at a loss. Give the agent landed cost, shipping, return rate and commission alongside the competitor feed, and make margin the decision variable.

Media budget follows the same logic. An agent optimising raw ROAS piles spend onto high-return products, because the return lands weeks after the click. Define the metric as net contribution after returns. An agent optimises whatever you hand it, quickly enough to make a bad metric expensive.

Storefront side: personalisation, on-site search and AI visibility

On the visitor side an agent does two jobs: it reads intent within the session and arranges the storefront around it, and it makes sure outside AI systems can parse what you sell.

Reading intent inside the session

A classic recommender looks at purchase history. An agent looks at the last ninety seconds. Someone three minutes into the size chart has a sizing question, not a product question. Instead of pushing similar items, the agent opens the fit helper, recalls the size from the previous order and quietly prevents a return. A visitor stalling on the shipping line gets delivery options, not another product.

On-site search is the weakest link in most stores. A box that returns nothing for "waterproof coat age 3" loses that session outright. Agent-driven search turns the query into a filter combination and, when nothing matches, shows the nearest alternative with a reason attached. Measure your empty-result rate for one week. That number usually says which area to hand over first.

Personalisation has a limit worth drawing early. An agent should not recite the visitor's history back at them. A reminder about last month's viewed item helps. A three-month-old search on the homepage feels like surveillance. The working rule: make the problem in front of the visitor easier to solve, nothing more.

Being readable to models, not just crawlers

A growing share of visitors arrive from a model's answer rather than a results page. If a model cannot read your product page, you cannot appear in that answer. A content agent keeps product data structured, fills missing attributes and publishes stock and price in machine-readable form. For the groundwork, the explainer on AEO and GEO visibility is a workable starting point.

Agents across logistics, shipping and returns

Everything after checkout decides most of your cost and most of your customer sentiment. A store in Turkey typically works with several carriers at once, sometimes two or three warehouses, plus marketplace fulfilment. Many decisions, very little time for each one.

Order routing and carrier choice

The moment an order lands, the logistics agent weighs the address, stock across warehouses, transit times per carrier in that district and the queue on the packing line. Routing a Kadıköy order to Bursa because Istanbul is backed up might cost one extra day and 40 TL. The agent prices that trade-off and declines the reroute above your threshold.

Carrier selection sits in the same calculation. Transit times for Yurtiçi, Aras and MNG vary by district, and the gap widens during campaign weeks. Rather than defaulting to one carrier, the agent reads the last thirty days of real delivery data for that address. Same price and two days faster wins. For the warehouse design underneath, the logistics and warehouse management guide gives a practical frame.

Returns and support

On a return request the agent reads the message first and the photograph second. If it identifies carrier damage, it files the claim, instructs the warehouse to ship a replacement and closes the loop in one reply. If the issue is sizing, it offers an exchange instead, which is cheaper than a refund and keeps the customer. Money moving out is the exception: the agent prepares the case, writes the justification and calculates the amount, then a person presses the button.

KVKK, IYS and the boundaries you set

Agents work fast. A badly scoped agent works just as fast. For a store operating in Turkey, writing the limits at the start is far cheaper than repairing them later.

Personal data

Under KVKK the decisive question is which fields the agent moves and where. Before sending a delivery address to a language model, ask whether the model needs it at all. District level is enough to estimate transit time. Full name and building number are not. Cover automated decision-making in your disclosure text, list your model provider among your data processors, and extend your retention policy to agent logs.

Access control

Key management here is a legal question as much as a technical one. The API key your agent uses should never be your own admin account. Give the agent its own user, grant only the endpoints it needs, and rotate the key quarterly. When something goes wrong, you need to separate what a person did from what an agent did without reconstructing it from memory.

Send its actions to a separate log stream too. When a customer disputes a price or a message they received, you want that answer in minutes rather than stitching the sequence together from three systems.

Permission-based marketing and marketplace rules

Any agent sending commercial electronic messages falls under IYS registration, so consent has to be checked before the message goes out. With no consent the agent shows an on-site notice when the visitor returns instead. Price and stock sync on Trendyol and Hepsiburada is natural agent work, but campaign rules there come from your seller contract. Cap the agent's marketplace pricing authority at what the platform permits.

How to get there in ninety days

Moving to agentic commerce need not start with replacing your platform. If the API side of your stack works, the first agent can be live in three weeks. Sequence matters more than speed.

Days 0 to 30: put the data in one place

An agent cannot work across scattered systems. Orders, stock, customers, shipping and ad performance have to be readable from one place. Add no new capability here, only connections. A catalogue with inconsistent product codes produces an inconsistent agent, so clean the stock codes once and do not revisit them.

Days 30 to 60: ship one agent

Give the first agent a job that is easy to measure and cheap to get wrong. Watching stock thresholds and pausing ads on sold-out items fits: the return shows up in lira within a week and the downside is bounded. Run it in shadow mode for thirty days, deciding without executing while you approve each call. Once approval passes 90 percent, switch the permission on.

Days 60 to 90: add the second and third

The second agent usually earns its keep in support, the third in content and pricing. Repeat the loop each time: narrow permission, shadow mode, log review, wider permission. Platform choice becomes decisive here, because a dashboard with no callable functions cannot host an agent however good it looks. The costly mistakes people make choosing a platform show the invoice you would otherwise pay in month four.

Tomorrow morning the job is small. Open your panel, list every product that hit zero stock in the last thirty days and add up how many days their ads stayed live. That figure is the budget for your first agent. To see which connections are ready, the integrations page lists what is already wired.

Frequently Asked Questions

How is agentic commerce different from a chatbot

A chatbot answers the question, an agent completes the task. The bot tells a customer where the parcel is. The agent notices the delay, files the carrier claim and messages the customer unprompted. The difference is not the model but the tools and permissions attached to it.

Is this too early for a small store

Below roughly 200 orders a month a full build is overkill, but one agent makes sense at any size. Small teams always have work nobody gets to: stock thresholds, ads left running on dead products, shipping questions unanswered overnight. The test is not revenue, it is how many repetitive tasks go undone.

What happens when an agent gets it wrong

It will. Every agent needs an undo path and a hard ceiling, so price changes roll back in one click and daily volume is capped. Shadow mode removes most of the risk, because you see the error pattern before anything reaches production.

Which tasks should stay with people

Money movement, contractual commitments and anything defining brand voice. Refunds, supplier agreements, homepage messaging and customer communication during an incident belong in that group. An agent prepares, drafts and calculates. A person presses send.

Do I have to replace my platform

Not necessarily, but the API side has to work. Without a readable and writable interface to order, stock and product data there is nothing for an agent to hold. The quickest check is one question to your technical team: can we update a product price without opening the admin panel.

What is MCP and why does it keep coming up

Model Context Protocol is a contract that lets language models call tools in external systems in a standard way. It spread quickly because it removes the need for a bespoke integration per piece of software. Define your functions properly once, and different models can use the same tools.

Does this create extra KVKK exposure

The exposure comes from the data you move, not the agent itself. Minimise the fields sent to the model, keep personal data inside your own systems unless there is a reason not to, and state automated decision-making in your disclosure text. Bring agent logs into your retention policy as well.

Who on the team should own this

The technical team can build it, but the owner should sit in operations. One person spending an hour a week reading agent logs and marking which decisions were right is enough. Leave the role empty and the agent becomes a black box nobody audits, switched off entirely after the first real mistake.