Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- What is a chatbot, in more depth?
- What are the types of chatbots?
- What does a "good" chatbot look like, and how is it measured?
- Common misconceptions about chatbots
- Misconception 1: "Chatbots are scripted decision trees that send you in circles"
- Misconception 2: "Modern chatbots hallucinate constantly"
- Misconception 3: "A chatbot replaces your support team"
- Misconception 4: "Every chatbot is an AI chatbot"
- What a chatbot is NOT: chatbot vs AI agent, copilot, and conversational AI
- How does My AskAI handle this?
- FAQs
Created time
Jun 29, 2026 04:16 PM
Title length (<60)
Author
Last optimised
Ecomm?
Image
Publish date
Jul 1, 2026
Video
Slug
what-is-a-chatbot
Featured
Type
Article
Ready to Publish
Ready to Publish
A chatbot is software that holds a conversation with a person in natural language, taking a typed (or spoken) message and replying with a relevant answer.
The "chatbots" most people remember were scripted dead-ends. Today the word covers everything from a button-menu to a system that resolves support tickets on its own, which is exactly why it misleads buyers.
That definition is true, and it is also where the confusion starts. The chatbots most people remember (the website pop-ups and "Press 1 for billing" menus of roughly 2015 to 2023) were scripted decision trees that funnelled you down a fixed path and usually left you typing "agent" to escape.
That experience is what the word "chatbot" still carries for most buyers. But that bad reputation is baggage the word picked up over a decade of clunky bots, and today's technology is a different animal.
Here is why this page exists. "Chatbot" has stretched to cover wildly different things: a button-driven FAQ widget is a chatbot, ChatGPT is a chatbot, and a system that reads a customer's order, issues the refund, and updates the ticket is also called a chatbot.
When one word covers all three, it stops telling you what you're actually getting. That's exactly why we (and most of the support industry) now say "AI agent" instead.
What is a chatbot, in more depth?
TL;DR: "Chatbot" covers four generations of technology, from scripted button-trees to systems that take action. The word names the conversation rather than the capability behind it.
A chatbot is any software you converse with in natural language. The trouble is that the label spans four very different generations of technology, and that's the main reason it confuses buyers.
The first chatbot, ELIZA, was built at MIT in 1966 and matched keywords to scripted responses to imitate a therapist (fun fact: people still got weirdly attached to it). For the next fifty years the idea barely changed. Rule-based bots followed hand-written scripts, and the "smarter" ones mapped what you typed to a fixed list of intents.
Then large language models arrived, and a chatbot could suddenly understand almost any phrasing and write a fresh answer rather than pick one from a menu (this is the shift that reset what we build). The newest generation goes further again and takes action: not just answering a question, but doing the thing the question was about.
I went and read what the top results actually say, and they capture the surface of all this but stop short of the part that matters to a buyer. IBM calls a chatbot "a computer program that simulates human conversation." AWS describes "a software application that can simulate human-like conversation." Oracle frames it around automating interactions.
All correct, all neutral, and (here's my gripe) none of them tell you that two products both called "chatbots" can differ by a decade of capability.
That's the thesis of this whole post: "chatbot" describes a conversation interface, not a capability level. It tells you the software talks to people, but nothing about whether it can actually resolve anything.
What are the types of chatbots?
TL;DR: The four types are rule-based, NLP/intent, generative (LLM), and agentic. The jump that matters is from answering questions to actually taking action.
There are four types of chatbot, and the leap I'd tell any buyer to focus on is the jump from "answers questions" to "gets the job done."
Type | Understands free text? | Grounded in your knowledge? | Takes action? | Typical use |
Rule-based / scripted | No (button/flow trees) | No | No | Simple FAQ deflection, menu routing |
NLP / intent-based | Partly (mapped intents) | Sometimes | Rarely | Intent classification, ticket routing |
Generative / LLM | Yes | Yes (via RAG) | No | Natural-language Q&A grounded in your docs |
Agentic | Yes | Yes | Yes (tools + workflows) | Resolves end-to-end: refunds, order lookups, account updates |
- Rule-based / scripted. The classic. You click buttons or pick from a menu, and the bot follows a flowchart someone drew in advance. It can't handle anything off the script, which is why these are the bots that earned the word its bad name.
- NLP / intent-based. A step up. Instead of buttons, you type, and the bot tries to match your message to one of a fixed set of "intents" it was trained to recognize. Handy for routing a ticket to the right queue, but still boxed in by the intents someone defined ahead of time.
- Generative / LLM. This is what most people now mean by an "AI chatbot." A large language model understands almost any phrasing and writes an original answer. On its own an LLM will happily make things up, so support-grade versions are grounded in your own content using RAG (retrieval-augmented generation): the model pulls the relevant help article or policy before it answers, so the reply is based on your facts.
- Agentic. The newest type, and the one to beat. It does everything the generative type does, then adds tools and decisioning, so it can take an action: look up an order, process a return, update a record. This is the line between a chatbot that talks and an agent that does.
So how does a modern support chatbot actually work, end to end? Here is the flow we see in practice:
- A customer sends a message.
- The model works out what they mean (any phrasing, no menu).
- It retrieves the relevant knowledge from your help center, docs, or past tickets.
- If the task needs an action and a tool is connected, it takes that action.
- It replies, and if it can't help or it senses frustration, it hands the conversation to a human.
What does a "good" chatbot look like, and how is it measured?
TL;DR: Judge a support chatbot on resolution rather than deflection. 65-80% autonomous resolution with steady or rising customer satisfaction is solid, and 80%+ is world-class.
For a support chatbot, judge it on resolution rather than deflection. A good one autonomously resolves 65-80% of tickets with customer satisfaction holding flat or rising, and 80%+ is world-class.
Why does that split matter? Because the easy metric to brag about (deflection, or "the ticket didn't reach a human") counts a customer who gave up and closed the chat as a success.
Resolution counts only the conversations where the customer's actual problem got solved. A chatbot that deflects 90% but resolves 30% is just a wall.
Here is the band we use, drawn from a first-party aggregate of AI resolution rates across roughly 55 vendors and 195 rated deployments, where the median lands near 70%:
Tier | Autonomous resolution | What it usually means |
World-class | 80%+ | Mature knowledge + live data + action tools; CSAT held or up |
Solid | 65-80% | Good knowledge coverage, clean handoff |
Average | 40-65% | Knowledge gaps or no action tools yet |
Early | 25-40% | Thin or stale knowledge, FAQ-only |
A few caveats on that aggregate, because careful benchmarking demands them (and I'd rather you trust the number than over-read it). It's a directional industry figure rather than an apples-to-apples comparison, since every vendor defines "resolution" slightly differently. It's also self-selected, because teams that publish numbers tend to be the ones doing well.
Real rollouts land across the whole band. TravelJoy reached 80% AI resolution after their previous Zendesk AI setup resolved 24% of tickets (same knowledge base, very different result), and Edel Optics hit 79% with 92% customer satisfaction, across multiple languages.
The number a chatbot can reach depends far more on the knowledge and tools behind it than on the brand name on the box.
Common misconceptions about chatbots
TL;DR: The three big myths are that chatbots are dumb scripts, that they hallucinate constantly, and that they replace your team. None of them hold up for modern systems.
Three myths do most of the damage when people talk about chatbots, and I hear all three on calls: that they're all dumb scripts, that they hallucinate constantly, and that they replace your support team. None of them survive contact with how modern systems actually work.
Misconception 1: "Chatbots are scripted decision trees that send you in circles"
True of the 2015-2023 generation, much less so for the LLM ones. The bots that built the word's reputation genuinely were rigid flowcharts, and if your only experience of "chatbot" is shouting "representative" at a phone tree, that frustration is earned (I've done plenty of shouting myself).
A generative or agentic chatbot has no script. It understands free text and reasons about your actual question. The bad reputation is real history; it's just history.
Misconception 2: "Modern chatbots hallucinate constantly"
Most wrong answers in deployed support AI come from stale knowledge rather than invention. The model faithfully repeats what your help center says, and your help center is out of date.
Grounding the model in current, well-maintained content (RAG) removes most of the risk, and a team-facing audit trail removes the rest. With us, you can ask Echo, the assistant inside the dashboard, why the agent gave any answer and which source it pulled from. The fix for "the bot said something wrong" is almost always "fix the underlying article", not "the AI is unreliable."
Misconception 3: "A chatbot replaces your support team"
It removes repetitive volume; it doesn't remove people. A well-run support chatbot handles the high-frequency, low-complexity questions (where's my order, how do I reset my password, what's your return policy) so your team spends its time on the hard, emotional, or genuinely novel cases.
The goal is a support team freed from answering the same question for the four-hundredth time, with people still very much in the building.
Misconception 4: "Every chatbot is an AI chatbot"
No, and this is the whole problem with the word. A button-tree pop-up with no AI in it anywhere is still, correctly, called a chatbot.
So when a vendor says "we have a chatbot", that statement is compatible with both a 2018-era flow builder and a system that resolves 80% of tickets. The label doesn't separate them, which is why you have to ask what's actually under it.
What a chatbot is NOT: chatbot vs AI agent, copilot, and conversational AI
TL;DR: A chatbot is the conversation. An AI agent adds decisions and action, a copilot helps a human, and conversational AI is the underlying technology rather than a product.
A chatbot is the conversation. An AI agent adds decisions and action on top (this is the distinction I care about most).
A copilot assists a human rather than replying to the customer, and "conversational AI" is the underlying technology rather than a product you buy.
Term | What it is | Difference from a chatbot |
Converses and takes action and decisions; runs multi-step workflows | The action and decisioning superset; a chatbot is just the conversational part | |
Drafts replies and does lookups for a human agent to send | The human sends; a chatbot replies to the customer directly | |
Conversational AI | The umbrella technology (understanding + conversation + generation) | A capability layer, not a product itself |
Virtual agent / assistant | A vendor synonym, often voice-inclusive | Mostly branding |
RAG | Grounds an answer in your knowledge | A technique used inside modern chatbots, not a product |
LLM | The model that understands and generates text | The engine; the chatbot is the product built around it |
This is where the "misleading term" thesis pays off, so watch what the category leaders actually call their products. Ada rebranded its "chatbot" as an "AI Agent", and Intercom's bot is now Fin, the AI agent.
Zendesk, Decagon, and we at My AskAI all lead with "agent", not "chatbot".
That migration is the industry trying to escape the taint the word "chatbot" still carries, and to signal that these systems make decisions and take action rather than reading from a script. If you want the deeper version of the modern, AI-powered end of this spectrum, we wrote a companion piece on what an AI chatbot is (this post is the wider view of the word itself).
How does My AskAI handle this?
TL;DR: My AskAI is an AI agent rather than a legacy chatbot. It works inside your helpdesk, grounds answers in your own knowledge, and takes action through Tasks and Tools.
We built My AskAI deliberately as an AI agent, not a "chatbot" in the legacy sense. Tasks replace decision-tree flows, so there's no flowchart to fall off.
It lives inside your existing helpdesk (Zendesk, Intercom, Freshdesk, Gorgias, or HubSpot) rather than bolting a separate widget onto your site. It grounds every answer in your own knowledge (help center, website, docs, and past resolved tickets) so replies are based on your facts, and it takes action through Tasks and Tools: looking up an order, processing a return, updating a record.
Whether it actions something on its own or drafts it for a human to approve is your choice, set per action (most teams start with propose-then-approve and open up autonomy as trust builds). When it can't help, it hands off cleanly to your team.
If your team ever needs to know why the agent said something, you can ask Echo (the assistant inside the dashboard) why the agent gave any answer and which knowledge source it used. That's how the rollouts above reached their numbers: TravelJoy at 80% resolution, and Edel Optics at 79% with 92% satisfaction. Pricing is usage-based at roughly $0.10 per ticket, and you can test everything (all features, unlimited tickets, no card) on a 30-day free trial.
FAQs
What is a chatbot, in one sentence?
A chatbot is software that holds a conversation with a person in natural language and replies with a relevant answer. As this post argues, the word now covers everything from a scripted button-menu to a system that resolves support tickets on its own.
What is an example of a chatbot?
Examples span the full range. A website's "Press 1 for billing" pop-up, ChatGPT, a WhatsApp ordering bot, and a support agent that reads your order and issues a refund are all chatbots. That range is precisely why the single word is so unhelpful.
How do chatbots work?
A modern chatbot takes your message, uses a language model to work out what you mean, retrieves relevant information from a knowledge source (this is RAG), optionally takes an action through a connected tool, then writes a reply. Older rule-based chatbots skip all of that and simply follow a pre-built script.
What are the different types of chatbots?
Four: rule-based/scripted (follows a flowchart), NLP/intent-based (matches you to fixed intents), generative/LLM (understands any phrasing and writes original answers), and agentic (also takes actions). The meaningful jump is from answering questions to actually completing tasks.
Is ChatGPT a chatbot?
Yes, ChatGPT is a generative chatbot. But it's a general-purpose one trained on the open web rather than grounded in your business's knowledge, so it works differently from a support chatbot that answers from your help center and acts on your systems.
What is a chatbot used for?
Common uses are customer support, lead capture, internal IT and HR help desks, and ecommerce ordering. In customer support specifically (the use we know best), the goal is to resolve the high-frequency repetitive questions automatically so human agents can focus on complex cases.
What's the difference between a chatbot and an AI agent?
A chatbot is the conversation; an AI agent adds decisioning and action. A chatbot answers a question, where an agent can answer it and then do something about it, like processing the return the question was asking for. Most modern "chatbots" worth buying are really agents, which is why we dropped the word.
What's the difference between a chatbot and a virtual assistant?
Mostly branding. "Virtual assistant" is a vendor synonym for a chatbot, sometimes implying voice support. There's no firm technical line between the two terms.
Are chatbots and conversational AI the same thing?
No, and we get this one a lot. Conversational AI is the underlying technology (the understanding, conversation, and generation layers), while a chatbot is a product built using that technology. You can have conversational AI inside many products, and a chatbot is one of them.
Why do people say "chatbot" is a bad word now?
Because a decade of scripted, dead-end bots gave the term a poor reputation, and because the word no longer separates a capable system from a frustrating one. The industry is shifting to "AI agent" partly to escape that history and partly to signal a genuinely different paradigm: one that makes decisions and takes action.
Do all chatbots use AI?
No. A rule-based chatbot is pure scripting with no AI at all. That's exactly why "chatbot" is misleading: the label applies equally to a system with no intelligence and one powered by a state-of-the-art language model.
Can a chatbot resolve a support ticket on its own?
Yes. A modern agentic chatbot can resolve a large share of tickets end-to-end, including ones that need an action like a refund or an account change, as long as it's grounded in good knowledge and connected to the right tools. Real rollouts commonly land in the 65-80% range (it's what we see across customers), with the best above 80%.
Do chatbots hallucinate?
A generative chatbot can produce a wrong answer, but in deployed support the usual cause is out-of-date knowledge rather than invention. Grounding the model in current content and reviewing answers with a team-facing audit trail keeps this in check.
What's a good resolution rate for a support chatbot?
65-80% autonomous resolution is solid and 80%+ is world-class, but only if customer satisfaction holds steady or improves. A high resolution rate paired with falling satisfaction usually means the bot is closing conversations rather than solving problems.
