Skip to content

AI agent

| Article: 3 min read | AI in e-commerce

An AI agent is a system built on a language model that decides for itself which steps and which tools to use to reach a goal, and then carries those steps out: it looks up data, opens pages, changes records. A chatbot replies to messages. An agent does tasks.

In plain words

A chatbot waits for a question and gives an answer. An agent gets a goal ("find three ways to ship to Lviv for under 150 UAH and book the best one") and builds the sequence of actions itself: it calls a calculator, the carrier's account, a table with your rules.

In its review of approaches, Anthropic separates two kinds of systems. In workflows, the model and the tools are tied together by code written in advance: a person defined the steps. In agents, the model itself directs the process and picks the tools, keeping control over how the task gets done (Anthropic, "Building effective agents", 2024, source). It's a useful distinction: many products called "agents" are really workflows.


Chatbot

Workflow

Agent

Who decides the steps

A person asks, the bot answers

Code written in advance

The model

Does it act

Usually not

Yes, by a fixed scheme

Yes, it chooses the actions

Predictability

High

High

Lower

Cost and speed

Low

Medium

Higher

What it looks like in e-commerce

  • Customer support. According to Anthropic, support is a good fit for agents: a conversation needs customer data, order history and a knowledge base, an agent can also take actions such as a refund or a ticket update, and success is easy to measure. More: AI in customer support.
  • A shopping agent. A program buys a product on a person's behalf. That is a separate topic: agentic commerce.
  • Internal assistants. Pull a report from the CRM and spreadsheets, draft a reply, reconcile stock. Connecting to systems often goes through MCP.
  • Development help. An agent writes code and runs tests. Human review matters most here (see vibe coding).

What the data says about readiness

  • Among surveyed developers (Stack Overflow, 2025), 14.1% use AI agents at work daily, 9% weekly, and 37.9% don't plan to use them. 87% agree they worry about the accuracy of information from agents, and 81% about data security and privacy (Stack Overflow, 2025).
  • In June 2025, Gartner predicted that over 40% of agentic AI projects would be cancelled by the end of 2027 because of cost and unclear value, and called most current projects early experiments. It also warned about "agent washing": chatbots and automation rebranded as agents, and estimated that only about 130 of thousands of vendors are genuinely agentic (Reuters via The Star, 2025). This is a forecast, not a count.

Risks to account for

Anthropic notes that agent autonomy means higher costs and latency, and errors can compound across many autonomous steps. The recommendation: test thoroughly in a sandbox, and set guardrails and stopping conditions, such as a maximum number of iterations (Anthropic).

In everyday terms: don't give an agent the right to take actions that can't be undone (deleting an order, refunding any amount, changing prices) without a limit and a human's confirmation.

Common mistakes

  • Using an agent where a workflow or a single model request would do.
  • Giving an agent broad permissions "just in case".
  • Not keeping an action log: after a mistake it's unclear what exactly the agent did.
  • Buying an "agent" by its name: check whether it chooses the steps itself or just runs a script.

FAQ

How is an AI agent different from a chatbot? A chatbot answers questions. An agent gets a goal and carries out a sequence of actions with tools: it looks up data, changes records, calls other services.

What is agentic AI in plain words? It is AI that is given a task rather than a question, and plans and carries out the steps itself.

Does a small online store need AI agents? Not always. For many tasks, a single request to a model with the right data or a ready-made workflow is enough. An agent makes sense where the steps can't be predicted in advance.

Is it safe to give an agent access to the CRM and payments? Only with minimal permissions, limits and a human's confirmation for irreversible actions. Most surveyed developers in 2025 (81%) worry about data security and privacy when working with agents.

What is agent washing? Gartner's term for products that are called agentic although they have no substantial agentic capabilities.

Sources

Want to check this on your store?

AI automation

An AI assistant checks the store every day, suggests fixes and makes them after your «Yes».

About the service →

Share

Related

  • Agentic commerce

    | Article: 5 min read

    Agentic commerce is shopping in which an AI agent acts on a person's behalf: it searches for and compares products, and within the permission it has been given it can also place the order and pay. For a store, this means part of the "visitors" will be…

  • Technical debt

    | Article: 6 min read

    Technical debt is the future cost of fixing and reworking that arises when software was built fast and without room for change. Like financial debt, it lets you move faster today, but with interest: every next change costs more and takes longer.

  • Vibe coding

    | Article: 4 min read

    Vibe coding is building programs where a person describes the task in ordinary language, AI writes the code, and the person accepts the result almost without checking it and goes by whether it "looks like it works", adding new wishes in the chat. For quickly…

  • How an online store team uses AI every day

    | Article: 2 min read

    AI in a store team's daily work is a set of small, repeating tasks in which a language model prepares a draft and a person checks it and decides. The value comes not from separate "smart" features but from simple habits assigned to specific people.