Conversational AI for customer service, explained

Conversational AI for customer service is software that reads a customer’s question in their own words and writes back an answer, in a chat or on the phone, at any hour. Today it’s usually a large language model that answers from your help content and hands the customer to a person when it can’t help. Most guides to it are written for contact centres, so they lead with phone menus, call routing and tools for rooms full of agents. If you answer support yourself, you need far less. You need a bot that understands a question however it’s asked, answers only from what you’ve written, replies in the customer’s language, asks when a question is vague and brings you the questions that need you.

Deacon is an AI customer support agent, a kind of conversational AI built for founders who answer support themselves. It answers your customers in seconds from your own docs, in whatever language they write in. When your docs don’t cover a question, it says so and asks for the visitor’s email instead of guessing. It also shows you what customers are stuck on and what they want next. There’s a free plan, and it needs no card.

What conversational AI does when a customer writes in

Take away the contact-centre extras and conversational AI in support does six jobs. Most have a page of their own on this site, linked below, so here’s what each one means for you.

It understands a question asked in plain words

Customers don’t use your words. They write “can I get my money back”, not “refund policy”, and they misspell your product’s name. Conversational AI works out what a question means instead of matching keywords, so a question nobody predicted can still reach the right answer. It also reads each message in light of the ones before it, so “and on the yearly plan?” makes sense as a follow-up. That’s the conversational part, and it’s what a search box or an FAQ page can’t do.

It answers from your own content

A language model on its own answers from what it picked up in training, which knows nothing about your refund policy. A support bot finds the passages in your help content that fit the question and writes its reply from those, which is called grounding. A good one also says it doesn’t know when those passages don’t cover the question. Grounding AI explains how that retrieval works, and why chatbots hallucinate covers the made-up answers grounding stops and the ones it doesn’t.

It replies in the customer’s language

A customer who writes in Spanish gets a Spanish reply, even when every page you’ve written is in English. The model writes in the customer’s language from whatever it found, so there’s no second set of docs to keep. The chat window’s own buttons and labels are a separate matter, and often stay in one language. Multilingual chatbot shows how to check replies in a language you don’t read.

It asks a question back when a question is vague

“It’s not working” could mean a dozen things. A good bot asks one short question, such as which calendar or which plan, instead of guessing or listing every fix it knows. It should also know when to stop, because a customer who has answered three questions and still has no answer won’t get one from a fourth.

It hands the customer to a person

Some questions need you, such as a refund, a bug or a sale that hangs on a missing feature. Conversational AI should recognise them, take a way to reply, such as an email address, and tell you. Someone who asks for a person should get one the first time they ask. Human handoff sets out when a chatbot should pass a chat on, and chatbot vs live chat weighs a bot against a person sitting in the chat.

It shows you what customers asked

Every question is a customer telling you what they needed. A tool that saves the conversations and groups similar questions shows you which docs are missing and which features people want, from everyone who asks instead of the few who email. Customer insights from the questions customers ask covers what to do with each kind.

Conversational AI, chatbots, generative AI and AI agents

The words overlap, and vendors use them loosely. In practice you’ll meet four kinds of support bot, and they differ most in what happens to a question nobody planned for.

A rule-based chatbot isn’t really conversational AI. It follows a script, and a customer who types something the script didn’t expect gets nowhere.

Intent-based conversational AI was the standard before large language models. In Google’s Dialogflow ES, for example, an intent “categorizes an end-user’s intention for one conversation turn”, and you give each one training phrases, “example phrases for what end-users might say”. It copes with many ways of asking a question you planned for. A question you didn’t plan for falls to the default fallback intent, which comes set up with replies such as “I didn’t get that. Can you say it again?”

Generative AI is the technology, and conversational AI is the job. Generative AI means models that write new text, images or code, such as the large language models behind ChatGPT. Conversational AI means software built to hold a conversation. Most conversational AI for customer service is now built on generative AI, which is why it copes with questions nobody planned for. The risk changes with it. A script can’t make anything up, and a language model can, so grounding matters more than it used to.

“AI agent” is used two ways. In support it often means the AI that answers customers, the way a support agent does, and that’s the sense in which Deacon is an AI customer support agent. In AI more broadly, an agent is a model that uses tools to act, and some support products sell exactly that. Intercom says its Fin agent handles “refunds, cancellations, account changes, and multi-step workflows” using live data from systems such as Stripe, Shopify and Salesforce. That’s what people mean by agentic AI, conversational AI that acts as well as answers. It needs your systems connected and rules for what it may do, which is worth the work once the same request arrives often enough to automate. Customer service automation for founders sets out what else to automate, stage by stage.

Conversation design, the craft of scripting what a bot says at each step, was most of the work with the first two kinds. With generative conversational AI it mostly becomes writing clear help pages and a few plain words about tone. How to write knowledge base articles is where that time pays off.

The contact-centre parts a team of one can skip

The pages that rank for conversational AI in customer service come mostly from contact-centre, CRM and cloud vendors, so they lead with what a contact centre needs. The analysts write for the same reader. Forrester’s Wave for conversational AI platforms for customer service, published on 16 April 2026, scores 14 platforms. Forrester’s post about it says more than 650 conversational AI vendors compete “in a market shaped by the needs of the roughly 15 million contact center agents working globally today”, and that the platforms must fit into “real-world contact center environments”, including “connecting to telephony and contact-center-as-a-service platforms”. Forrester’s public pages don’t list the 14, whose places the vendors announce themselves. It’s a shortlist for a company that runs a contact centre.

Here’s what those guides cover, and whether you need it when you answer support yourself.

• Phone menus and voice bots. Interactive voice response, or IVR, answers your support line and routes callers by what they say or press. If customers don’t phone you, skip it.
• Agent assist. It suggests replies to a human agent during a live chat or call, which pays off across dozens of agents answering all day. For you, the bigger saving is not being in the conversation at all.
• Routing, queues and shift planning. These send each conversation to the right agent and plan who works when. With one or two of you answering, there’s nobody to route to.
• Every channel in one inbox. WhatsApp, SMS, social messages and email in one place matter once customers write to you there. If they find you through your website, the chat on it is where their questions arrive.
• Account data from a CRM. Large platforms read a customer’s record to tailor an answer or act on it. A lighter version does for most small teams, where your site tells the bot which plan the visitor is on.

What you can’t skip are the six jobs above. A contact centre needs them on a phone line with a hundred agents behind it, and you need them in a chat on your website with you behind it. If you’re weighing a contact-centre suite anyway, Deacon vs NiCE and Deacon vs Five9 set out who each one suits.

What conversational AI can’t do

Conversational AI is good at questions whose answer is written down. These are where it stops, and a good tool says so instead of trying.

• It can’t see a customer’s own account. Their invoice, their usage and why their card failed aren’t in your help pages. Unless it’s connected to your billing and checks who’s asking, the honest reply is that it can’t see that, with a way to reach you.
• It can’t decide. Refunds, exceptions and discounts are your call. An AI agent can carry out a refund you’ve written rules for, but no bot should be making up policy.
• It can’t know what nobody wrote down. A grounded bot says it doesn’t know, and an ungrounded one guesses, which is how bots end up promising refunds that don’t exist.
• It can still be wrong. An out-of-date page gets repeated faithfully, and a search can pick the wrong passage, so every conversation should be saved for you to read.
• It doesn’t take phone calls unless it’s built for them. Voice is a separate product, with speech recognition and a phone line. Deacon answers in a chat on your website and takes no calls.

One support chat with Deacon, turn by turn

Here’s what those jobs look like in a real product. Acme is a made-up booking app with Deacon on its site, and Lena is a made-up customer. Acme’s help pages cover calendar sync, and like any help pages they say nothing about one customer’s bill. Each behaviour below is what Deacon does. Its wording would differ, because Deacon writes each of these replies fresh.

1. A vague question gets one question back

Lena: the calendar sync isn’t working

Deacon: Sorry it’s giving you trouble. Which calendar are you connecting to Acme?

Acme’s help pages have separate steps for Google and Outlook, so the right answer turns on one detail Lena hasn’t given. Deacon asks for that one thing instead of guessing or listing both. If three questions back still don’t get her an answer, it stops asking, sums up what she told it in one sentence and asks for her email so you can take it from there.

2. The answer comes from Acme’s own help page

Lena: Outlook. New bookings aren’t showing up in it.

Deacon: Go to Settings › Calendars, disconnect Outlook and connect it again, and allow Acme to edit your calendar when Microsoft asks. New bookings should then appear within 15 minutes.

Deacon read Lena’s reply in light of her first message, searched Acme’s help content and wrote the answer from the passages it found on connecting Outlook, in seconds. A link to that page usually sits under the reply. Had Lena written in German, the reply would have come in German, from the same English page.

3. A question about her own bill is declined

Lena: Also, why was I charged twice in August?

Deacon: You’d like to know why you were charged twice in August. I can’t see your account or your invoices, so leave your email below and the team will be in touch.

Under the reply, the chat asks “Want the team to follow up?” above a box for her email. Deacon always declines a question about the visitor’s own account, invoice or data, whatever your help pages say, because it can’t see them. It never guesses at a double charge. Later in a conversation, a decline opens by summing up what the visitor has said, so you don’t have to ask again. The decline doesn’t end the chat either, and Deacon would still answer Lena’s next question about Acme.

4. She leaves her email

Lena types her address and sends it, and the box gives way to “The team will get back to you.” In your dashboard, the conversation shows as unread.

5. She asks for a person

Lena: Can I talk to an actual person about this?

Deacon: Yes, the team will be in touch at the email you left.

A request for a person is taken at its word the first time it’s made. Lena has already left her address in this conversation, so Deacon doesn’t ask for it again. In your dashboard, this turn is tagged Wants a person.

6. You get one email, and reply in her chat

Three minutes after Lena left her address, the owners on your team get an email with her last message and the page she was on. Her request for a person came within those three minutes, so it arrives as one email, not two. If one of you already has the dashboard open, no email goes out, because the dashboard has already told you.

You open the conversation and type your reply. Sending it takes the conversation from Deacon, which stops answering in it. If Lena still has the chat open, your reply appears there under your first name and your company’s, such as “Priya from Acme”, and the box she types in changes from “Ask AI” to “Reply to the team”. The chat stays open to your reply for twenty quiet minutes, and after that, or if she has left, your reply goes to the address she gave. When you’re done, press Hand back to Deacon.

The whole exchange is saved word for word, with the source each answer named. If other customers have asked about Outlook sync too, Topics groups their questions with Lena’s and counts the people who asked, which tells you the reconnect step belongs somewhere easier to find. And had Lena’s first question been one Acme’s pages don’t cover, Deacon would have said it didn’t know and asked for her email, and her question would wait in your dashboard for your answer.

Why Deacon fits a founder who answers support

Deacon is built for the person in step 6, a founder with no support team and no helpdesk. It does the six jobs above in a chat on your website, with nothing to build first. There are no flows, decision trees or intents. You add your website, files, pasted text or written answers, try it in the Playground, and paste one script tag on your site or hand the prompt to Claude Code, Cursor or Codex.

It answers in seconds, at any hour, in the language each customer writes in, and only from what your content says. When your content doesn’t cover a question, it says so and asks for the visitor’s email. You’re emailed when they leave it or ask for a person, and your reply from the dashboard appears in their chat while it’s open. That works on every plan, with no helpdesk in between.

It gets better the more you answer. Every question your content didn’t cover waits in Conversations with an Answer this button. Write the answer once and press Save & check, and Deacon shows you the reply it now gives and how many other open questions your answer covers. Every answer you write becomes part of what Deacon knows, so it needs you less over time.

It also shows you what customers need. Every conversation is saved and searchable by what visitors asked. Topics group similar product questions, count the people who asked and label most of them as a how-to, a bug report, a feature ask and so on. Analytics › Widget shows how many people wrote in, the share answered and the languages they wrote in. You hear from everyone who asks, including customers who’d never have emailed you.

Every feature is on every plan, the free one included. Free gives you 50 answers a month, with no card and no end date. Starter is $50 a month, or $40 a month billed yearly, for 1,000 answers a month, around 500 conversations. Nothing is billed past a plan’s answers, no plan upgrades itself, and every plan answers with the same model.

Questions about conversational AI for customer service

Is conversational AI the same as a chatbot?

Not quite. A chatbot is any bot that chats, including one that only offers a script of buttons. Conversational AI is the kind that understands questions in plain words and answers in its own words. A conversational AI chatbot is where the two meet, and it’s what most people now mean by an AI chatbot.

Which is the best conversational AI for customer service?

The one that fits who answers your customers and where they write to you. A contact centre with a phone line needs a platform from a list like Forrester’s. A founder answering support alone needs a chat on the website that hands over to them. The best AI chatbots for customer service ranks nine for small teams, with prices and the plan where a person can take over.

Can conversational AI replace a support team?

Not all of it. It can take the questions your docs answer, at any hour and in the customer’s own language, and pass the rest to a person. It can’t make judgement calls, and it can’t see accounts it isn’t connected to. For a founder, that means you mostly hear about the questions that need you.

How much does conversational AI for customer service cost?

Platforms for contact centres are mostly priced through a sales call. Tools for small teams bill by the answer, the credit, the resolution or the conversation, and several have free plans. Deacon’s free plan gives you 50 answers a month, and Starter is $50 a month, or $40 a month billed yearly, for 1,000. How much does an AI chatbot cost? prices one month of questions each way.

Can Deacon answer phone calls or WhatsApp messages?

No. Deacon answers in a chat on your website, and you reply to customers from its dashboard. It has no phone or voice agent, no WhatsApp, Instagram or SMS channel and no support inbox for email.

Can Deacon take actions, such as refunds or plan changes?

No. Deacon answers questions and hands conversations to you, and it takes no actions in other systems. It always declines questions about a visitor’s own account, invoice or data. Your site can pass Deacon details about the visitor, such as their plan, and Deacon takes them into account when it answers.