Conversational AI is software that talks with people in natural language on channels like chat, WhatsApp and the phone. Generative AI is the technology that writes new text, speech or images, and agentic AI is a system that works out the steps towards a goal and takes actions in other systems to complete them.
They are not competing products. Each word describes a different part of the same system, and only one of them tells you whether it can do anything.
One customer, three systems
Take one message of a kind every ISP sees often. This is an illustration, not a real customer.
"My bill has gone up by twelve pounds this month and nobody told me. What is this?"
The scripted bot
The older chatbot is conversational in the loosest sense. It matches "bill" to a branch somebody drew in advance and replies with a menu: view your bill, payment options, speak to an agent.
The customer picks "view your bill" and gets a link to the account portal. Their question was why the bill went up, and that question has gone unanswered. Most people type "agent" at this point, which is the bot's real job as far as they are concerned.
The generative bot
Put a language model behind the same chat window and the reply gets much better. It understands the question and writes a clear answer about the common reasons a bill changes: a promotional discount ending, an annual price rise, an add-on, a one-off charge.
All of that might be accurate, and none of it is about this customer. The bot cannot see the account. It can only guess, politely, and suggest they check their bill or contact the team. The customer leaves better informed and no further on.
The agent
An agent has the same conversational front end and the same ability to write well, plus access to the billing system and permission to do some things in it.
It verifies the customer, reads the account and finds that a twelve-month discount ended this month. It explains that in one sentence. It also notices the customer was never sent the end-of-discount notice, which your policy says should have gone out. Your rules allow a goodwill credit for that, so it raises the credit for the difference, confirms it in the chat, logs the reason on the account, and offers to look at plans that bring the monthly price back down.
Same question. Only the last one answered it for this customer.
The difference in plain terms
Conversational describes the interface. Is the customer talking in their own words, on a channel they already use? A scripted bot and an agent can both be conversational. For more on that, see what conversational AI is.
Generative describes how the words are produced. Instead of picking a reply from a list, the system writes one. It made bots far better at understanding and answering, but writing is all it does on its own.
Agentic describes what the system can do. It can choose steps towards a goal and use tools to carry them out: look something up, change a record, book an appointment, raise a ticket. That is the line drawn in chatbot vs AI agent, and it is the one that changes your resolution numbers.
What agentic needs in practice
The model is the easy part. An agent is only as useful as what sits around it.
Access to your systems. Billing, CRM, helpdesk, network status, engineer booking. Each one needs an integration the agent can call, which is usually where most of the build time goes. More on that in can conversational AI integrate with my CRM.
Permissions. Decide what it may do and how far. Reading an account is one thing. Issuing a credit is another, and it needs a ceiling: up to a set amount, for set reasons, once per customer in a set period. Anything beyond that goes to a person.
Guardrails. Identity checks before it reveals or changes anything. Rules it cannot talk itself out of. A record of every action it took and why, so your team can audit it.
Handover to a person. Some conversations need a human: a complaint, a vulnerable customer, anything outside its permissions. A good agent hands over with the customer already verified and the issue written up, so nobody repeats themselves. See what happens when an AI agent does not know the answer.
Where the labels get abused
Each of these words has been stretched to sell something.
"Conversational AI" gets used for decision trees with a text box on the front. If it gives you a menu when you type a sentence, it is a scripted bot.
"Generative AI" gets sold as if writing well were the same as helping. A model that writes a perfect paragraph about why bills change has still not looked at the bill. That gap is covered in whether a general AI assistant is good enough for customer service.
"Agentic" is the newest, so it is the most stretched. Products get the label because they can send an email or fill in a form on the vendor's own platform, with no connection to yours.
A simple test: ask the vendor to name the systems the agent writes to, what it is allowed to change, and where the limits are. If the answer is about how natural it sounds, it is not an agent.
You can see the difference for yourself on our own agent.
