Insights · 16 Nov 2026

What is an AI agent? Chatbot, assistant and agent compared

A chatbot answers. An assistant helps while you decide. An agent acts for you. What each term means, what sits inside an agent, what it changes for a Shopify store, and what Singapore's guidance asks of anyone who uses one.

Every software vendor now sells an "AI agent". Some of these products are chatbots with a new label. Others can log in to your tools, send emails and place orders. The difference matters, because it changes what can go wrong and who is responsible.

The short answer

A chatbot answers. An assistant helps, and you decide. An agent acts on your behalf.

OpenAI defines agents as systems that independently accomplish tasks for you. Anthropic draws a similar line. It separates workflows, where the steps are fixed in advance in code, from agents, where the model itself decides which steps to take and which tools to use. Google Cloud ranks the three by autonomy. Bots follow pre-set rules. Assistants respond to requests and leave the decisions to the user. Agents pursue a goal with more independence than the other two.

Singapore's IMDA notes that there is no single agreed definition. The common thread is some degree of independent planning, decision-making and action.

Chatbot, assistant and agent side by side

ChatbotAssistantAgent
What it doesAnswers questions, often from a script or a knowledge baseDrafts, summarises and suggests when you askPlans and carries out a task across several steps and tools
Who decidesRules written in advanceYou, at each stepThe model, within limits you set
Typical exampleThe FAQ window on a shop websiteA writing or research helper inside an appA system that reads new enquiries, checks the calendar and drafts replies for approval

The labels are not strict. Many products sit between two columns. The same product can also move from one column to the next on the day you connect it to your accounts.

One reason for the overlap: chatbot describes an interface, and agent describes a capability. One product can be both.

A simple test: ask who chooses the next step. If it is you, you have an assistant. If it is a script, a chatbot. If it is the model, and it can use tools to change something outside the chat, you have an agent, even when it asks for your approval first.

What is inside an agent

IMDA's Model AI Governance Framework for Agentic AI lists the parts in plain terms.

  • A model. The language model that reasons and plans. IMDA calls it the brain of the agent.
  • Instructions. Text that sets the agent's role and its limits.
  • Memory. Information the agent can store and recall.
  • Tools. Connections that let it act: write to a file or a database, send a message, make a transaction.
  • Controls. Access limits, guardrails and logs that bound what the agent may do and record what it did.

Tools are what turn a chat window into an agent. Controls are what make it fit for work.

A current example: agents that pay

On 13 November 2026, OpenAI introduced ChatGPT Wallet. It lets an agent in ChatGPT complete a purchase for its user. The user adds a card once, reviews the items and the total, then approves each purchase with a check such as Face ID or a fingerprint. The business receives a payment token, not the real card details.

Two points are worth noting. The launch is limited to ChatGPT Pro subscribers aged 18 and older in the United States, with US-issued cards. And every purchase needs an approval. OpenAI says it expects users to delegate purchases within set limits later. That is not the case today.

This is a fair picture of where agents stand in late 2026. They can act. For anything that costs money, this one still asks first.

A concrete case: a D2C brand on Shopify

Take a skincare brand in Singapore that sells direct to consumers from a Shopify store. A shopper asks an assistant for a light, reef-safe sunscreen for humid weather, under S$40, delivered this week. Agents already touch that store in three ways.

  • They read it. Any public product page can be read by AI assistants and their agents. HUMAN Security found that in October 2026, 85% of AI agent requests went to product and search pages, and under 2% to checkout. One agent, Meta's Muse, made 59% of those requests. Agents compare facts first: ingredients, sizes, price, stock, delivery times and the return policy.
  • Shopify sends them your catalogue. Shopify's agentic storefronts share structured product data with AI channels through Shopify Catalog. It is active by default for eligible stores, managed in Sales channels > Agentic.
  • Some can sell for you. On a few channels the shopper can buy without leaving the chat. On the others, the agent sends the shopper to your own checkout.

What each channel does today, for a store selling to buyers in Singapore:

AI channelHow it gets your productsWhere the purchase happensFor buyers in Singapore
Meta AI and MuseShopify CatalogDirect checkout on Meta surfaces, including Muse where available. On by default for eligible storesAvailable: Singapore is one of the 11 eligible buyer countries
Google AI Mode and GeminiYour Google Merchant Center feedDirect checkout for buyers in the US, Canada and Australia; elsewhere, your storeProducts can appear; buyers finish on your store
ChatGPTShopify CatalogYour store's checkout, opened from ChatGPTOnly if your store also sells to US buyers
Microsoft CopilotShopify CatalogDirect checkout, on by default, shown only to buyers in the USOnly if your store also sells to US buyers

The use case is not "a robot buys your products". It is a shortlist. The agent picks two or three products that match the question, and the brand with complete, accurate facts has a better chance of being on it.

Five things a Shopify brand can check this week:

  1. Open Sales channels > Agentic and see which channels are on, and whether direct checkout is on for each.
  2. Complete your policies. Terms of service, privacy policy and return and refund policy are required for every AI channel.
  3. Write product facts, not slogans. Shopify's own advice: materials, dimensions, price and availability, clear variant names, and images that match the product.
  4. Fill in the knowledge base. Shopify keeps store facts such as returns, shipping times and brand values there, so agents can answer customer questions.
  5. Keep your Google feed in sync through the Google & YouTube sales channel, so AI Mode and Gemini see current prices and stock.

To keep a product out of these channels, set its status to Unlisted.

What Singapore says

Singapore has written guidance for both organisations and individuals.

For organisations, IMDA launched its framework for agentic AI in January 2026 and updated it in May with case studies. It rests on four ideas: assess and limit the risks upfront, keep humans meaningfully accountable, put technical controls in place, and help end users act responsibly.

For individuals, CSA, GovTech and IMDA issued a joint advisory on personal AI agents on 6 November 2026. It is written for consumers. Its five steps also make a sound starting list for a small business:

  1. Limit what the agent can access.
  2. Stay in control of important actions.
  3. Find out how your information is used and stored.
  4. Protect yourself from scammers, including instructions hidden in web pages or emails.
  5. Think about the impact on other people.

The advisory is also clear on responsibility. You remain responsible for what your agent does on your behalf.

Do you need an agent?

Often, not yet. Anthropic's advice to builders is to look for the simplest solution and to add complexity only when it clearly improves the outcome. Agents usually cost more and respond more slowly than a single prompt, and their errors can compound.

Three questions help:

  • Is the task repeated often, with clear inputs and a clear result?
  • Can you name the actions that need a human approval?
  • Can you read a log of what the agent did?

If all three answers are yes, an agent may be worth a trial on one task. If not, an assistant and a checklist are enough for now.

Where to start

Pick one task. Give the agent read-only access first. Keep a person on every action that sends, spends or deletes. OpenAI's own guide makes the same point: sensitive or irreversible actions should go to a human until the agent has earned confidence.

If you want help choosing that first task, get in touch.

Sources

Checked on 16 November 2026.

Updated 16 November 2026: replaced the simple test, which measured human approval and not autonomy, noted that chatbot is an interface and agent a capability, clarified where direct checkout is available on Meta and Microsoft Copilot, added the share of agent requests made by Muse, and added two Shopify sources.

Start with a conversation

Tell us where you are stuck. We reply within two working days.