AI agent vs chatbot vs assistant vs concierge: what is the difference?

The category labels in AI products are a mess in 2026. The same word means three different things to three different audiences. A "chatbot" to a banking ops team means a tree of decision rules with maybe an LLM at the leaves. A "chatbot" to a startup founder means GPT-in-a-box. An "assistant" to Apple means Siri. An "assistant" to OpenAI means a configurable instruction-tuned profile inside ChatGPT. An "agent" can mean almost anything depending on who is selling.

This post is the cleanest categorical taxonomy we have been able to build, in plain English, with examples. It is the version we wish someone had handed us in 2023.

Chatbot

A chatbot is a decision tree, optionally with LLM-generated text at the leaves. The structure is rule-based: if user asked X, respond Y; if user said Z, escalate to human. Most "chatbots" sold to enterprises in 2024-2026 are still this category, just with friendlier copy.

What it is good for: high-volume, narrow-scope, deterministic flows. A bank's "what is my balance" path. An airline's "rebook my flight" path. A SaaS support page that answers the top 20 questions. The chatbot wins when you want predictability over depth.

What it is bad at: anything outside the tree. You cannot ask a chatbot to do a thing it was not pre-built for. The development cost is in building the tree; the maintenance cost is in keeping the tree current as the underlying business changes.

Assistant

An assistant is a chat surface backed by a language model, with optional instructions/profile and optional retrieval. ChatGPT is the dominant example. Pi, Claude, Gemini, and a long tail of vertical "AI assistants for X" all fit this shape.

What it is good for: long-form conversation, drafting, brainstorming, research. The assistant is an extraordinarily good thinking-companion when the goal is to think out loud, get a draft, or refine a piece of writing.

What it is bad at: doing things in the world. The assistant has no native concept of "calendar," "inbox," "Stripe account," or "the customer who called yesterday." It can describe what should be done; it does not do it. Many products try to bridge this with "actions" or "plugins" but the user is still inside the chat surface, copying outputs into the real tools.

Concierge

A concierge is the hybrid: human operators backed by AI assistance. The customer thinks they are talking to a person (and often they are, partially). Behind the scenes the AI drafts the responses, surfaces the relevant context, and the human approves before sending.

What it is good for: high-trust, high-touch services where a wrong answer is expensive. Wealth management, complex healthcare scheduling, white-glove travel planning. The concierge is also the right answer for early-stage AI products that cannot yet be trusted on their own; the human-in-the-loop is your safety net during the trust-building period.

What it is bad at: scale economics. Concierge services have a real human cost per customer interaction. The unit economics work for high-margin work and break for high-volume, low-margin work. Many startups call themselves "AI concierges" when they really mean "we have AI inside but a human always reviews"; that is fine, but it is a different category of business than a pure software product.

Agent

An agent is the category we are interested in, the one most consumer AI products mean when they say "AI." An agent is a system that takes goal-directed action on a user's behalf using one or more language-model loops plus real tools (real APIs, real software, real money). The loop is what makes it an agent rather than a single-call assistant. The tools are what make the agent useful rather than entertaining.

What it is good for: multi-step real-world action. "Book my haircut," "send the invoice and follow up if unpaid," "answer my phone in my voice and book the job." The agent picks up a goal, does the work, reports back. The user did not have to know the steps.

What it is bad at: long-form chat. The agent's job is to act, not to converse. A user who wants to think out loud for 90 minutes is going to be happier in an assistant. Many people end up with both: the assistant for thinking, the agent for doing.

How to tell which one a product is in 60 seconds

Three questions:

Does it have rules I can see? If the demo shows a "flow editor" with branches and conditions, you are in chatbot territory. If the demo shows a chat window with no flow scaffolding, you are in assistant or agent territory.

Does it close the loop? Ask the demo to do a task end-to-end ("send this email to john@example.com," "book me a meeting with X"). If the response is a draft you have to copy, it is an assistant. If the response is "done," and the email actually went, it is an agent.

Is there a human in the middle? If the company describes a "team" of operators backing the AI, it is a concierge. If the product is software-only, it is an assistant or agent.

When each one is the right answer

Where Ajeris sits, plainly

Ajeris is in the agent category. We do not run rule trees. We do not stop at drafts. We do not put humans in the middle. Each user gets their own per-user agent that listens on the channels they already use, has tools to act in the world (Gmail, Calendar, Plaid, Spotify, your CRM, your bank), and closes the loop. When you ask the agent to send an email, the email goes.

We use ChatGPT and Claude for thinking-out-loud work the same way our users do. The categories complement, they do not compete.

Try it

The fastest way to feel the difference between an assistant and an agent is to ask both for the same thing. Open ChatGPT and ask "send a note to my mum saying I love her." Then text Ajeris with the same request. Compare what happens.