Diesen Artikel gibt es noch nicht auf Deutsch. Zu den deutschen Artikeln
AI agent vs LLM is a question of who the AI works for and what it can reach. An LLM (large language model) is a model that reads text and writes text from what it learned in training, for the person typing to it. An AI agent is a program built around an LLM that adds your content, your tools, your rules and a place to work, so it can answer your customers and act for them. If you use ChatGPT for your own emails, a chat assistant on its own is enough. For customers asking on your website or WhatsApp at 9 PM, you need an agent.
AI agent vs LLM: what is the difference?
No, an LLM is not the same as an AI agent, whichever model sits inside: the difference is not how clever the model is but what surrounds it.
A language model on its own is a prediction engine for words. It knows what was in its training data up to a cutoff date. It keeps nothing between two separate requests unless the app around it sends the earlier messages again. And it cannot press a button anywhere: whatever it "does" is text. Ask it whether you have a free slot on Thursday and it can only guess, or tell you to check.
An agent puts that model in a loop. It reads the message, decides what it needs to answer it (a page from your knowledge base, a lookup in your booking system, a missing detail from the customer), gets it, looks at the result and only then replies. The model chooses the next step; the agent gives it steps to choose from. The table spells it out, with a third column for the case this page is about: an agent answering on your own website.
Can ChatGPT answer your customers on its own?
Not on your website, and not from your systems. Paste the FAQ into ChatGPT, ask it a customer's question and get a good answer. That proves the model can write. It does not put the model in front of your customers.
Four things are missing in that setup. The answer goes to you, in your chat window, not to the visitor on your site or in WhatsApp. It knows what you pasted, not today's free appointments, stock or order status. Nothing connects it to your booking or order system for that visitor. And when a customer wants a person, nobody on your team is told. Whether ChatGPT itself counts as an agent is a separate question, answered in our guide to AI chat agents.
So for an owner, "AI agent vs ChatGPT" is really a question of who sits on each side of the chat. A chat assistant serves the person typing. A customer-facing agent talks to strangers in your name, all day, and stays inside what you allow. The two also work together: you can manage your agent from ChatGPT through an MCP connection, as our ChatGPT MCP guide shows.
Do I need an AI agent?
Maybe not. If the AI only helps you with your own work, a language model in a chat app is the right tool: drafting emails, writing product copy, summarizing a long document, rewording a price list. Nothing there needs your customers, your systems or a handover. You need an agent when at least one of these is true:
- Customers ask questions outside your working hours, and an evening message that waits until morning is a lost booking or sale.
- The right answer depends on your own data: your prices, your services, your opening hours, your stock.
- A request ends in an action: a booking, an order lookup, a quote request, a newsletter signup.
- Some conversations need a person, and you want them handed over instead of guessed at.
- Customers reach you on more than one channel, and you want the same answers on the website, WhatsApp, Messenger and Instagram.
If none of these applies (a few questions a week, all answered by email within your hours), keep the assistant you have and save the money. If two or more apply, a plain model will keep sending customers to "please call us", which is the answer they came to your website to avoid.
One customer question, answered by a model and by an agent
The quickest way to see what an agent adds is to send both the same question. Take Northside Physio, a fictional physiotherapy practice that keeps its appointments in booking software with a public API. At 9 PM a visitor writes on the website: "Do you have anything free Thursday afternoon?" Setup one is a language model given the practice's FAQ text. Setup two is an agent with the same FAQ plus a booking skill.
The model alone does its best with what it has. It cannot see the calendar, so it replies "Please call us to check." Polite, correct, and no help at 9 PM.
The agent works through the request. The skill asks the booking system for Thursday's free slots and gets two: 2:30 PM and 4:00 PM. The agent offers both and asks for the name and email the booking needs. When the visitor picks 4:00 PM, it sums up the appointment and asks "Shall I book it?" Only after the yes does it create the booking, and it confirms the time the booking system returned, not the one it expected.
Then the visitor asks whether their insurance covers the treatment, which the practice decides case by case. The agent says what the knowledge base supports and offers once to pass the question to the team. If a teammate is switched online, the visitor can also press "Talk to a human" in the website chat, and the colleague takes over the same conversation with its history.
What are the functions of an AI agent?
An AI agent has six jobs around the model, and each one closes a gap the model leaves open. Take any of them away and you are back to a model that writes well and knows little about your business.
Knowledge
Answers from your content: website pages, PDF and Word files, Q&A pairs, Google Drive, OneDrive, Notion and product catalogs.
Skills
Acts in your tools: checks free slots, looks up orders, creates contacts or payment links, through ready-made skills or your own API.
Rules
Follows the behaviors you write: tone, length, topics it never discusses, promises it never makes. They apply on every channel.
Memory
Keeps the conversation in view, so "the second one" or "4 PM works" makes sense without the customer repeating the question.
Handover
Knows when to stop: in the website chat a person takes over while someone is online; otherwise it offers to forward the request to your team.
Channels
One agent for your website, a chat link, WhatsApp, Messenger and Instagram, with the same knowledge and skills everywhere.
Knowledge matters most for trust. An agent that looks up your content before it answers a question answers from your content, not from guesses about your business, and you fix a weak answer by adding the missing page or Q&A pair. How sources are read and kept current is on the knowledge base page.
Build the agent around the model with Mo
Ultimo Bots is the AI agent platform where you describe in plain words what the agent should do, and Mo, the builder in the AI Studio, builds it. The model is already in place; what you add is your knowledge and the skills for your business. For Northside Physio, the owner would type something like this:
When a patient asks for a free appointment, check the open slots in our booking software and offer two. Ask for name and email, sum up the appointment and book it only after the patient says yes. Here are the API docs: <link>Mo reads the API documentation, asks for the API key in a secure field, builds the skill and tests it. You then chat with the agent in the Test tab as a visitor would, and "What happened?" under each answer shows which skill it used and what it looked up. Bookings you make there are real, so use a slot you can cancel. Ready-made skills cover Wix, Calendly, Cal.com, Stripe, HubSpot, Mailchimp and WooCommerce order status, and remote MCP servers connect much like an API.
What does an AI agent cost compared to an LLM subscription?
A chat assistant subscription is priced per person who uses it; an agent is priced by the answers it gives your customers. A personal plan pays for one person's questions. Developers who call a model directly pay for the amount of text it processes. An agent on your website pays for every customer answer, because every answer runs the model.
With Ultimo Bots that cost is counted in AI credits. As of November 17, 2026 the plans are:
- Smart: $29 a month per agent, or $19 a month billed yearly, with 500 AI credits.
- Boost: $59 a month, or $39 a month billed yearly, with 2,000 AI credits.
- Ultimo: $249 a month, or $199 a month billed yearly, with 10,000 AI credits.
A standard answer uses 1 AI credit, and Mo's building work draws on the same balance. Replies your team types in the live chat cost no credits, and teammates are free, so the bill grows with customer conversations, not with headcount. The pricing page has every plan detail, and what an AI agent costs works through real monthly volumes.
Start with the questions your customers already ask
Your next 10 minutes: open the last two weeks of customer emails and messages and sort them into two piles: questions your website already answers, and requests where you had to look something up or do something. The first pile is what an agent's knowledge base covers on day one. The biggest item in the second pile is the first skill to describe to Mo. If the second pile is empty, the language model you already use may be all you need.
Questions and answers
Is ChatGPT an agent or LLM?
It is an assistant built on large language models. When it only writes a reply, that is the model at work; when it uses tools to finish a task, it behaves like an agent. Either way it serves the person typing in it. An agent on your website serves your visitors instead, from your content and with your tools.
What is the difference between an AI model and an AI agent?
An AI model is the trained engine: it takes an input and produces an output, such as text, an image or a forecast. An LLM is one kind of AI model. An AI agent is software that uses one or more models to reach a goal: it decides the next step, calls tools, checks the result and replies. The model supplies language and judgment; the agent supplies knowledge, tools, rules and channels.
Can an AI agent work without an LLM?
Yes, but it is a different kind of agent. Rule-based agents, such as a thermostat or a scripted phone menu, follow fixed if-then rules and need no language model. They fail as soon as a customer phrases something nobody planned for. The agents businesses now put on their websites use an LLM to read free text and choose the next step.
What are the 5 types of AI agents?
The usual list is simple reflex agents (react to the current input), model-based reflex agents (keep track of what they cannot see), goal-based agents (plan toward a goal), utility-based agents (weigh which outcome is best) and learning agents (improve from feedback), as IBM lists them, read November 17, 2026. A website agent is goal-based within one conversation; AI agent vs agentic AI covers how much autonomy that should mean.
Where does MCP fit between an LLM and an agent?
MCP (Model Context Protocol) is an open standard for connecting AI applications to tools and data. The LLM decides, the agent runs the loop, and MCP is one way the agent reaches a tool. An Ultimo Bots skill can call a remote MCP server.
Does Ultimo Bots train its models on my customers' chats?
No: its help center states that visitors' conversations are not sold to third parties and not used to train Ultimo Bots' own models. The agent gets better because you improve it: add a page or a Q&A pair when a question goes unanswered, or ask Mo to change a behavior or a skill.




