Chatbot vs Conversational AI: One Is Inside The Other

The three terms, defined properly

Two of these are categories and one is a product type. Reading them in order makes the hierarchy obvious.

Conversational AI

Conversational AI is the umbrella category for any software that interprets human language and responds in human language, whether the exchange happens in text, in speech, or both.

It is a field rather than a product, in the same way that "computer vision" is a field. Nobody buys conversational AI as such. You buy something built with it: a chat widget, a phone bot, a voice assistant, an in-car command system, a triage bot inside a clinical app.

Because the term describes a category, vendors use it freely and it carries almost no information about capability on its own. A scripted keyword-matching bot from 2016 and a system that resolves a billing dispute over the phone can both be described as conversational AI without anyone technically lying. When a vendor uses the phrase, the useful follow-up question is which modalities it actually supports and what it does when the request is not one it was configured for.

Examples: Amazon Alexa, Contact-centre voice bots, NiCE Cognigy, Kore.ai, Google Assistant

Chatbot

A chatbot is a conversational AI application with a text interface: the user types a message into a chat window and the system replies in text.

The defining feature is the modality, not the intelligence. A chatbot can be a rule-based decision tree or a retrieval-augmented language model, and both are chatbots. What makes it a chatbot is that the conversation is written on both sides, in a window on your site, in Slack, in Messenger, or in a mobile app.

This is the box matram.ai sits in, and it is worth being blunt about the boundary. We answer typed questions from your own content across a website widget, Slack, Messenger and a handful of helpdesks. There is no voice, no phone number, no speech recognition. Whether that is a limitation or simply the correct scope depends entirely on where your customers are trying to reach you.

Examples: matram.ai, Intercom Fin, Zendesk AI agents, Tidio Lyro

Intelligent virtual assistant (IVA)

An intelligent virtual assistant is a conversational AI application that spans multiple modalities and can take actions in back-end systems, not just answer questions.

IVA is the term the contact-centre industry uses for the systems that replace phone queues. The distinguishing features are usually voice plus text on the same conversation state, integrations that let it reset a password or reschedule a delivery rather than describing how, and handover into an agent desktop with the transcript attached.

The vocabulary is unstable here. Zendesk describes its AI agents as "autonomous systems designed to understand and autonomously resolve complex issues on any channel" that, "unlike traditional chatbots", "can reason across multi-step requests". Other vendors would call exactly that an IVA. The label matters less than the two questions underneath it: does it handle voice, and can it change anything in your systems. We go further into the assistant end of this spectrum in chatbot vs virtual assistant.

Examples: NiCE Cognigy, Kore.ai, Ada, Amazon Alexa

The hierarchy: where a chatbot sits inside conversational AI

Conversational AI divides most cleanly by modality, because modality is what determines the engineering, the cost and the vendor list.

Text-based conversational AI

Chat windows on websites, in-app messengers, chatbots inside Slack or Teams, social channels such as Messenger, and general-purpose chat interfaces. The input arrives as clean text, which removes an entire class of problem: no accents, no background noise, no transcription errors to recover from. This is why text products are cheaper to build, cheaper to run and quicker to deploy than their voice equivalents, and why almost every self-serve conversational product is a text one.

Chatbots live here. So does the chat interface you use with a general model, though a hosted assistant and a chatbot trained on your own content are not the same thing.

Voice-based conversational AI

Smart speakers, in-car assistants, phone-line virtual agents and contact-centre IVAs. Everything in the text stack still applies, plus two extra stages: speech has to be converted to text on the way in and back to speech on the way out, both in real time, both while the caller is waiting. Voice systems also have to handle interruption, silence, hold music, DTMF keypresses and the fact that a caller cannot see a list of options.

That added surface area is the honest explanation for the price gap later on this page. A voice deployment is a telephony project as well as an AI project.

Multi-modal conversational AI

One conversation that moves between channels without starting over: a caller who is sent an SMS link mid-call and continues in a browser, or a support thread that begins in chat and escalates to a call with the transcript carried across. Multi-modal is the hardest of the three, because the difficulty is not the language model at all. It is shared conversation state, identity resolution and integration into every system that holds a piece of the customer record.

Read the hierarchy top down and the buying decision falls out of it. If everything you need happens in a text box, you are shopping in the smallest and cheapest branch, and you should be suspicious of anyone quoting you for the whole tree. Our types of chatbots guide breaks the text branch down further.

When a chatbot is not the answer

We sell one of these, so read this section sceptically. Here is the honest ramp, from the point where you need nothing at all to the point where a chat widget is the wrong shape entirely.

Answering by hand is genuinely fine

Plenty of sites should not have a chat widget on them. If the questions that arrive fit inside one person's morning, and they arrive by email, adding a bot adds a thing to configure and a thing to keep watching. It also puts software between you and the moment you learn what confuses your customers, which is worth more in the early years than the minutes it saves. Nobody regrets having answered their first hundred customers personally.

Where the friction starts

The signal is repetition, and it usually appears in your own drafts before it appears in a queue. You keep a saved reply. You have sent it several times this week. Somewhere in your help centre is a page saying the same thing, and people are not finding it. That is a search problem wearing a support problem's clothes, and it is exactly what the text branch is built for. It is also, worth noting, not what a conversational AI platform is for.

Where it becomes a liability

It turns into a liability when the unanswered question has a buyer attached to it. Evaluation traffic does not wait around. Someone comparing you against alternatives late at night reads silence as an answer, and you never see the conversation, because it never started. At this stage the coverage gap is costing revenue rather than time, and the fix is the cheap branch of the tree rather than the whole tree.

The case that breaks it

And then there is the case where none of this helps. If the queue that genuinely hurts is on a phone line, a chat widget does not shrink it, because the people waiting are not looking at a screen. Adding text automation to a business whose customers ring is a way of being busy in the wrong branch. The same is true of expecting a chatbot to change something inside an account: retrieval finds pages, and a page is not an order record. Both of those are category problems, and no amount of configuration inside a chatbot platform reaches them.

What is actually under the hood

Five acronyms do most of the work in this field, and vendor marketing tends to list all of them regardless of what a product contains. Here is what each one honestly is.

NLP: natural language processing

The broad discipline covering everything computers do with human language: tokenising, parsing, classification, translation, summarisation, generation. NLP is the parent term, so "powered by NLP" tells you close to nothing. Spell-check is NLP.

NLU: natural language understanding

The part concerned with extracting meaning from input: working out what the user wants (the intent) and which details they supplied (the entities). In classical chatbot architecture, NLU was a separate trained component. You listed intents such as "track_order", gave each one dozens of example phrasings, and the model classified incoming messages into one of them. Every question outside the intent list fell through to a fallback.

NLG: natural language generation

Producing the reply. In classical systems this was mostly templating: a fixed sentence with slots filled from the entities. It was predictable and it was rigid, which is exactly the trade-off described in the rule-based versus AI comparison.

LLMs: large language models

The change since 2022 is that a single large model now performs the NLU and NLG steps together, without an intent list. You give it the user's message plus relevant passages from your own content, and it composes an answer. That is retrieval-augmented generation, and it is what most current AI chatbots do, including ours. The practical consequences are that setup no longer means enumerating intents, and that coverage is now limited by the quality of your source content rather than by how many phrasings someone remembered to type in.

It is worth knowing that older intent-based NLU has not disappeared. Voice systems in particular still use intent classification for the parts of a call that must behave identically every time, with a language model handling the open-ended stretches.

ASR and TTS: speech recognition and speech synthesis

Automatic speech recognition turns audio into text; text-to-speech turns the reply back into audio. These are the two components that separate a voice product from a text one, and they are also where most of the perceived quality of a voice bot comes from. Latency matters more than accuracy at the margin, because a caller reads a one-second pause as the system being broken.

A chatbot platform contains no ASR and no TTS. That is not a feature gap to be fixed in a future release; it is the definition of the product category.

Chatbot platform vs conversational AI platform

Both phrases appear on vendor homepages and they describe genuinely different products. The reliable tells are voice, back-end actions and whether you can see a price without talking to anyone.

Gartner's June 2025 press release coined the term "agent washing" for vendors rebranding assistants, RPA and chatbots as agentic AI without the underlying capability, and found that only around 130 of the thousands of vendors claiming agentic AI were genuine. The same inflation happens one rung down with "conversational AI platform". If a product has no voice path and no way to write to your systems, it is a chatbot platform with a bigger word on the homepage.

What each type of platform actually costs

None of the three major conversational AI platforms below publishes pricing on its own website. As of 20 July 2026, cognigy.com/pricing and kore.ai/pricing both return HTTP 404, and ada.cx/pricing shows a consultation link with no rates. Their AWS Marketplace listings do carry public prices, which is where these figures come from.

Two warnings about that table. First, the units are not equivalent: a conversation contains several messages, an "outcome" on Fin's pricing page means a resolution, procedure handoff or disqualification, and a Kore.ai session is defined per module. Treat the last column as order of magnitude, not as a like-for-like rate. Second, the enterprise Cognigy line is a private SaaS licence covering a whole contact centre, so its low unit cost only exists at volumes almost nobody reaches.

The pattern is still clear. A conversational AI platform for a mid-sized deployment starts in the tens of thousands a year and involves a sales cycle before you see a number. A chatbot platform is a monthly subscription you can read off a page. If you want to put your own volumes against those rates, the chatbot ROI calculator does the arithmetic with your inputs rather than ours, and the Intercom pricing breakdown covers the per-outcome model in detail.

What buying in the wrong branch costs you

The price gap between the branches is in the table above. The costs worth worrying about are the ones that never appear on a quote.

Buy a platform sized for a contact centre when all your customers do is type, and the licence is the least of it. You have signed up for an implementation project: workshops, a discovery phase, an internal owner, integration meetings with people whose systems you do not need. Most of that effort goes into capability you will never switch on. And it is spent at the precise moment when the fastest available win was to start answering the questions people are already asking on your site.

Buying too small costs you differently. If phone really is where your customers wait, a text chatbot does not reduce that queue slightly. It reduces it by nothing, and you now run two systems that know different things about the same person. The customer who chats and then calls starts again from the beginning, which is the exact experience every vendor in this market claims to be fixing. Your team ends up reconciling the gap by hand, which is a job nobody agreed to take on.

Then there is the internal cost of having been visibly wrong. Somebody championed the purchase. When it does not land, the next proposal to automate anything meets a colder room, and the organisation quietly learns that this whole category overpromises. That lesson is expensive, because it makes the correct purchase harder to get approved a year later than it would have been today. So the question to ask before signing is not whether the platform is capable. Capable of clearing which of your queues, and how would you prove that inside the first month?

Which one you need

The deciding question is not how sophisticated you want to be. It is which channel your customers use when they have a problem.

Being direct about our own position: matram.ai is a chatbot platform, text only. We do not do voice, telephony, IVR or multi-modal handover, and we are not going to describe ourselves as conversational AI infrastructure to win a comparison. If your actual problem is a call centre, the vendors to evaluate are NiCE Cognigy, Kore.ai and Ada, and you should expect the prices in the table above. We are not on that shortlist and would be wasting your time if we pretended to be.

Where we do belong is the text branch: a SaaS product whose customers ask the same forty questions about setup and billing, or a small business that cannot staff live chat all day. If you are already inside a helpdesk and weighing its native bot against a dedicated one, the Zendesk alternative comparison and the wider alternatives index are the honest place to start.

Frequently asked questions

Sources

• Gartner press release, 25 June 2025 - The term "agent washing" and the finding that only around 130 of thousands of vendors claiming agentic AI were genuine.
• Anthropic, "Building effective agents", 19 December 2024 - The definitions of agents and workflows used to separate agency from conversational interface.
• Zendesk, AI agents - Zendesk's own definition of AI agents as autonomous systems that resolve complex issues on any channel.
• AWS Marketplace vendor listings, accessed 20 July 2026 - Ada Strategic at $33,000 per 12 months for 60,000 conversations; NiCE Cognigy Basic 5K pm at $43,080 per 12 months; NiCE Cognigy enterprise private SaaS at $1,000,000 per 12 months; Kore.ai per-session rates of $0.015 to $0.20.
• Ada pricing page, accessed 20 July 2026 - Ada publishes no rates on its own site, only a consultation link.
• Cognigy pricing page, accessed 20 July 2026 - The page returns HTTP 404, so Cognigy publishes no pricing on its own site.
• Kore.ai pricing page, accessed 20 July 2026 - The page returns HTTP 404, so Kore.ai publishes no pricing on its own site.
• Fin pricing page, accessed 20 July 2026 - $0.99 per outcome, a 50-outcome monthly minimum, and one outcome billed per conversation, where an outcome is a resolution, procedure handoff or disqualification.