Chatbot Examples: 12 Real Bots, Verified Live

Twelve chatbot examples at a glance

Ordered roughly by how easy each is to see for yourself. Some need an account, one is inside a government app, and two are products you would buy rather than bots you would visit.

Two entries are worth flagging as different in kind. Fin, the helpdesk AI agent sold by Fin (the company formerly known as Intercom), and matram.ai are products you buy and point at your own content, not deployments of somebody else's brand. They are here because they are the closest thing to a live demo you can run yourself in an afternoon.

Customer service chatbot examples

These are the ones people mean when they search for customer service chatbot examples. Each is a support deployment at real volume, and each has published something about how it went.

IKEA Billie

IKEA's own customer service pages describe Billie as available 24/7 to answer common questions before, during and after a purchase, reached by clicking the yellow chat bubble that follows you around ikea.com. The name is a nod to the Billy bookcase, which is a small thing that does real work: customers remember it, and a named bot is easier to escalate away from than an anonymous one.

The design detail worth copying is placement. Billie is not parked behind a Contact Us link at the bottom of the site. It is on product pages and order pages, where the question actually occurs. Ingka Group, the largest IKEA franchisee, has also written publicly about what it did with the capacity Billie freed up, retraining call centre staff into remote design advisors rather than cutting them. That is the part most vendors never show you.

Klarna's AI assistant

Klarna's own February 2024 press release said the assistant handled two thirds of its customer service chats in its first month and did work equivalent to about 700 full-time agents. Those are Klarna's numbers about Klarna, and worth reading as such. What makes this the single most useful example on the page is what happened next: in 2025 Klarna publicly said it had leaned too hard on cost, that service quality had dropped, and that customers would always have the option of reaching a person.

So the lesson is not the headline number. It is that a company with one of the most aggressive AI support deployments in the world landed on a hybrid, with the bot on routine questions and humans on everything else. If you are planning a rollout, plan for that endpoint rather than for the press release.

Vodafone TOBi

TOBi handles billing, device and SIM questions inside the My Vodafone app and on the web, and Vodafone's newsroom describes a generative upgrade, SuperTOBi, rolling out across multiple countries. The thing to notice is that it is one assistant reused across app, web and messaging channels rather than three separate bots with three separate sets of answers drifting apart. Telecom support is high volume and highly repetitive, which is the profile where this works best.

Amtrak Julie

Julie is the oldest example here and the most instructive about scope. Amtrak's own page describes her as a virtual travel assistant who knows the content of Amtrak.com and answers typed questions about schedules, fares and bookings, and the same persona answers the phone line. Julie has lasted because the job is narrow and the content behind her is stable. Nobody asked her to be a general assistant, and she is still running while flashier bots from the same era are gone.

GOV.UK Chat

The UK Government Digital Service launched GOV.UK Chat in the GOV.UK app in May 2026 after two public pilots. What makes it the best-documented example on this list is that GDS published the accuracy work: more than 10,000 users asking around 26,000 questions across the pilots, with accuracy scores moving from 76% at the start to 90% by the end, and the service built on Amazon Bedrock using Anthropic's Claude models. Web rollout is still described as future work.

Copy the sequence, not the technology. Pilot in a channel where you can measure, publish what accuracy actually is, and expand only once the number justifies it. Very few private deployments have the nerve to show their evaluation results, which is precisely why this one is useful.

When none of these examples is the answer for you

Twelve working deployments make a chatbot look inevitable. It isn't. Before you copy any of them, here are four situations where the honest recommendation is to build something else, ordered from the clear-cut to the genuinely arguable.

You are looking at Erica and you sell candles

The bank examples are the most copied and the least copyable. Erica and Eno answer questions about one customer's money by querying core systems, with all the authentication and audit that implies. That is an engineering programme, not a subscription. If there is no account data sitting behind your bot, promising the Erica experience sets an expectation your content cannot meet, and the gap between the two is where the disappointment lives.

Nobody has written the answers down yet

Every website example here works by answering from published content. IKEA's Billie and GOV.UK Chat are both sitting on top of material somebody wrote and somebody maintains. Take that away and there is nothing to ground against. A retrieval bot pointed at a thin help centre does not politely summarise what is missing. It improvises, or it declines all day, and both of those are worse than the search box you already have.

The job is to persuade, not to answer

Look at what Lemonade's AI Maya is actually doing. It is not working through support tickets. It is running a purchase flow one question at a time, and the conversational format is there because a long form loses people. If your real problem is that visitors don't convert, a chat widget bolted onto the same confusing page will not fix it. Fix the page. A bot in front of a confusing offer just hands people a new way to ask why it is confusing.

You could do the job with a form, and you know it

Here is the borderline one. A lot of what people want a chatbot for is intake: collect the person's details and their problem, then route it somewhere. A form does that today, with no source curation and no risk of an invented answer.

But the case for a bot was never the collecting. It is what happens before it. Duolingo's Video Call holds a conversation, and Shopify's Sidekick acts inside your admin, and neither of those is a form with a personality on top. So the test is simple. If the whole job is to gather a name and an email address, you have bought a chat-shaped form and you will maintain it forever. If the bot can answer the question that made somebody open the form in the first place, and only falls back to collecting details when it genuinely cannot help, it is earning its keep. The question is not whether a chatbot is possible. It is whether the answering part is real.

Banking and insurance chatbot examples

Financial services runs the largest chatbot deployments in the world, and they share one property: the bot answers questions about your account, not about a help article.

Bank of America Erica

Erica lives inside the Bank of America mobile app and answers questions about balances, transactions, spending patterns and where to find things in the app. Bank of America's newsroom reported in August 2025 that Erica had passed three billion client interactions since its 2018 launch, and in March 2026 put the running total above three billion in the context of its wider digital figures.

What a visitor sees is the important part. Erica answers "how much did I spend on groceries last month", which is a query against the customer's own transaction data, not a lookup in a knowledge base. That is a different and harder product than a website FAQ bot, and it is why bank bots are a bad template for a small business. If your bot has no account data to reason over, do not promise the Erica experience.

Capital One Eno

Eno does a similar job over a channel most people ignore: plain SMS. Capital One's product pages describe asking Eno about balances, available credit, minimum payments, due dates and recent transactions by text message, in the app, or signed in on the web, and getting virtual card numbers for online shopping. The lesson is channel choice. Text message has no install step and no login friction, so for a narrow set of high-frequency questions it beats any widget.

Lemonade AI Maya and AI Jim

Lemonade is the best insurance chatbot example because the bot is the product surface rather than a support add-on. AI Maya runs the quote and purchase flow on lemonade.com, asking one question at a time in conversation instead of presenting a long form. AI Jim takes claims.

The numbers here come from Lemonade's own annual report on Form 10-K for the year ended 31 December 2025, which is about as good as a source gets: AI Jim took the first notice of loss without human intervention 96% of the time, and roughly 55% of claims were fully automated from start to finish. The design lesson is in the other 45%. AI Jim is not authorised to settle every claim, and it routes anything outside its authority, or anything that trips a concern, to a human claims expert. The handoff is a defined rule, not a fallback for when the bot gets confused.

Website and in-product chatbot examples

These are the chatbot website examples that are easiest to imitate, because the bot answers from published content and admin data rather than from a core banking system.

Shopify Sidekick

Sidekick sits inside the Shopify admin and helps merchants with setup, product copy, images and questions about their own store. Shopify's page for it says plainly that "Sidekick is included with your Shopify plan. Features and usage limits vary by plan", and that it "has direct access to your Shopify data, understands commerce workflows, and takes action in your admin".

That last clause is the line between an assistant that answers and one that does. Taking action in the admin is a meaningfully different risk profile from answering a question, and it is worth being precise about which one you are building. Most SMB deployments should answer first and act later.

Duolingo Video Call with Lily

Duolingo Max subscribers can call an AI character called Lily and hold a spoken conversation in the language they are learning, with the app supporting English, Spanish, French, German, Italian, Portuguese, Japanese and Korean, and Duolingo's own blog describing transcripts, adaptive difficulty and memory of previous calls. It is the clearest example on this page of a bot whose job is not deflection at all.

The design choice worth stealing is that mistakes carry no penalty. Duolingo says explicitly that the point is low-pressure practice. If your bot's purpose is to get someone to try something, removing the cost of getting it wrong is more important than accuracy.

Fin, from the company formerly known as Intercom

Fin is the reference implementation of the helpdesk AI agent, and its pricing page is unusually direct about what it charges for: $0.99 per resolution with a fifty-outcome monthly minimum, one outcome billed per conversation, and $9.99 per qualification. Salesforce agreed in June 2026 to acquire the company for around $3.6bn, and the helpdesk itself is still sold under the Intercom name.

Look at it as an example even if you never buy it, because outcome pricing forces a definition. Somebody had to decide what counts as a resolved conversation, and that definition is the thing your own project will argue about for a month. Our Intercom pricing breakdown goes through how that maths lands at different volumes.

matram.ai

Our own bot belongs on this list as an example of the small-business end of the category, not the top of it. You crawl a URL, import a sitemap, upload PDFs or DOCX files up to 10MB, or connect Notion, Google Drive, Confluence, Zendesk and similar sources. The widget is one line of JavaScript, and the same bot runs in Slack, Messenger, Crisp, Freshchat, Zendesk, Zoho SalesIQ and Google Chat. WhatsApp and HubSpot are not live yet.

The design decision we would point at is citation. Every answer links the page it came from, and strict mode keeps the bot inside your approved content, so a wrong answer is traceable to a document you can go and fix. The honest limits: we are a newer platform with a smaller integration library than Intercom or Zendesk, there is no free tier, we hold no SOC 2 or ISO 27001 certification, and the product is text only. If you need any of those, buy something else.

Chatbot conversation examples: what good exchanges look like

Screenshots go stale faster than the bots do, so here are the four exchanges that separate a good deployment from a bad one, written out as scripts you can compare your own transcripts against.

1. A grounded answer that shows its work

Visitor: do you ship to Ireland and how long does it take

Bot: Yes. Standard delivery to Ireland is 3 to 5 working days
     and costs EUR 6.95, free over EUR 60.
     Source: Shipping and Delivery

     Want me to check whether an item is in stock?

Two facts, one source link, one forward step. The citation is not decoration. It is the mechanism that lets a support lead audit a hundred conversations in an hour and find the page that is out of date.

2. An honest no

Visitor: can I use my student discount on top of the sale price

Bot: I do not have anything in our help content that covers
     combining the student discount with sale pricing, so I do
     not want to guess on this one.

     I can pass this to the team now. What is your email?

This is the exchange the Air Canada and Cursor failures below were both missing. A bot that will say "I do not know" is worth more than one that is right slightly more often but confident every time.

3. A handoff that keeps the context

Bot: I have passed this to the billing team with your order
     number (#48812) and everything we have discussed.
     They reply on weekdays within a few hours.

     You will get the answer at sam@example.com. Nothing else
     needed from you.

Named team, the context already attached, a realistic time window, and an explicit statement that the visitor is done. The commonest escalation failure is not the routing, it is making a person retype what they just typed.

4. A qualification that earns the question

Visitor: how much is it for a team of 30

Bot: Every plan includes unlimited seats, so 30 people costs
     the same as 3. The plan is chosen on chatbots, messages
     and data sources instead.

     Roughly how many support conversations a month? I will
     tell you which tier that lands in.

Answer first, then ask. Bots that demand an email before saying anything useful get abandoned, and the leads they do capture are the ones who had already decided. If you want more of these, the chatbot script examples page has full flows.

What copying the wrong example costs

The documented failures below cost their companies headlines. The quieter cost is what happens when you pick the wrong example to imitate, and it lands nowhere near the invoice.

A quarter spent building the wrong shape

Take Erica as your model and you will spend a quarter discovering that the chat interface was never the interesting part. Take Klarna's launch announcement as your model and you will spend that same quarter building for a volume you do not have, then quietly rebuilding towards the hybrid Klarna itself settled on. Neither of those is a wasted subscription. It is a wasted planning cycle, and planning cycles are the scarcer thing.

The credibility you don't get back

Somebody in your company argued for this. They stood in a meeting and said it would work, probably with a slide carrying one of the logos in the table above. If the bot then invents a policy, the damage is not really to the bot. It is to that person's next proposal, and to everybody's appetite for the one after. Cursor's users cancelled over a rule that did not exist. The internal version is quieter and lasts longer: the widget stays switched on, nobody trusts it, and support goes back to answering the easy questions by hand just to be sure.

The documentation you didn't write instead

This is the one people miss. The work that makes a content-grounded bot good is writing the pages that are missing and retiring the ones that lie. That work pays for itself anyway. It improves search and it shortens tickets, whether or not a bot ever reads it. So if the effort goes into vendor selection and channel configuration, and the bot then underperforms because the content underneath it was thin, you have paid twice: once for the tool, and once by not doing the thing that would have worked either way.

The customer who just leaves

Air Canada's chatbot produced a tribunal ruling, which is at least legible. Most bad answers produce nothing anybody can see. The customer reads a confident wrong paragraph, believes it, and either acts on it or gives up on you. And that is why citations matter more than any other single feature on this page. They turn an invisible loss into a traceable one. A cited wrong answer points at a page you can go and fix this afternoon. An uncited one is a churn number with no story attached.

What the good examples share, and what the failures share

Twelve examples across banking, retail, telecom, government and language learning have very little technology in common. The patterns are all product decisions.

Notice that four of the five failures are grounding and guardrail problems, not model problems. A better model would have invented the Air Canada policy more fluently. What prevents these is restricting the bot to approved source material, citing what it used, and building the route to a human before you build anything else. Our chatbot best practices guide covers the launch checklist in detail.

Nine famous chatbot examples we removed

These appear on most chatbot example lists published this year. None of them survived checking. If you find them on another list, you now know how old that list really is.

• H&M on Kik. Built for Kik's Bot Shop marketplace. Kik announced in 2019 that it was shutting the messenger down and cutting staff, the app was later sold, and the bot marketplace did not come back. There is no live H&M Kik bot to try.
• Sephora's Messenger and Kik bots. Every source describing them dates from 2016 and 2017, and Sephora's own site refused all our verification requests, so nothing current confirms a live bot. Cut for lack of evidence rather than declared dead.
• eBay ShopBot. No current eBay page describes it. Separately, eBay ended support over Facebook and X for US users in March 2026 while moving to AI self-service, which is a different thing.
• Starbucks "My Starbucks Barista". Extensively covered in 2017, unmentioned by Starbucks since. Unverifiable today, so cut.
• Domino's "Dom" as a live consumer text bot. The current deployments we could find are regional and voice or WhatsApp based, not the Messenger ordering bot the old lists describe.
• Babylon Health's symptom checker. Babylon filed for Chapter 7 bankruptcy in the US in August 2023, sold its UK business, and wound down. It is the example most often still listed as a healthcare chatbot success.
• Woebot. The consumer app was retired on 30 June 2025. Its founder told STAT the cost of meeting FDA requirements, and regulatory uncertainty around language models, drove the decision. The company pivoted to selling through providers and payers.
• McDonald's drive-thru order taking with IBM. The test ended and the technology was switched off in all restaurants by 26 July 2024. It is a voice example anyway, but it turns up on text chatbot lists constantly.
• Expedia "Romie". Announced in 2024 and still described as an alpha inside Expedia's EG Labs with no launch date. Real, but not something you can go and use.

Two of these deserve more than a line. Babylon and Woebot were both health chatbots, both were widely praised, and both are gone for reasons that had nothing to do with whether the bot answered well. Regulatory cost and business model killed them. If you are evaluating a chatbot for a regulated field, the durability question is about the vendor, not the answers. That is also why we do not sell into clinical use: matram.ai has no HIPAA compliance and no business associate agreement, and a chatbot vendor that will not say so plainly is a warning sign.

The general lesson for anyone building a list, or reading one, is that consumer messaging bots from the 2016 to 2018 wave almost all died with the platforms that hosted them, while bots owned by the company on the company's own property mostly survived. Every one of the twelve verified examples above lives on a website, an app or a helpdesk the brand controls. That is not a coincidence.

Frequently asked questions

Sources

• Bank of America newsroom, "A Decade of AI Innovation: Erica Surpasses 3 Billion Client Interactions" (Aug 2025) - Erica is live, and the three billion interactions figure, as Bank of America's own published claim.
• Capital One, "Eno, your Capital One assistant" - Eno is a current Capital One product available over SMS, the mobile app and signed-in web, and what it answers.
• Lemonade, Inc. Annual Report on Form 10-K for the year ended 31 December 2025 (SEC) - AI Maya and AI Jim as current products, the 96% first-notice-of-loss figure and roughly 55% of claims fully automated at 31 Dec 2025.
• Klarna press release, "Klarna AI assistant handles two-thirds of customer service chats in its first month" (Feb 2024) - Klarna's own first-month claims about chat share and agent-equivalent work, attributed as the company's figures.
• CX Dive, "Klarna changes its AI tune and again recruits humans for customer service" - Klarna's 2025 reversal, the quality admission and the commitment that customers can always reach a human.
• IKEA customer service, "Chat with IKEA's chatbot Billie" - Billie is live, available 24/7, reached via the yellow chat bubble across the IKEA site.
• Ingka Group newsroom, "AI and Remote Selling bring IKEA design expertise to the many" - Ingka's account of Billie's role and the retraining of call centre staff into remote design advisors.
• Vodafone newsroom, "Meet SuperTOBi, Vodafone's new generative AI virtual assistant" - TOBi is a current Vodafone assistant and SuperTOBi is rolling out across multiple markets and channels.
• Amtrak, "Meet Julie: Your Virtual Amtrak Travel Assistant" - Julie is a current Amtrak virtual assistant answering typed questions about Amtrak.com content.
• Government Digital Service blog, "Answers in seconds, 24/7: GOV.UK Chat launches in the GOV.UK app" (14 May 2026) - GOV.UK Chat launch date, app-only availability, and that it is built on Amazon Bedrock with Anthropic's Claude models.
• Inside GOV.UK, "5 things we learned testing GOV.UK Chat" (16 March 2026) - The pilot scale (10,000+ users, ~26,000 questions) and accuracy moving from 76% to 90% across the trials.
• Duolingo blog, "Duolingo's Video Call Offers Realistic Conversation Practice" - Video Call with Lily is a live Duolingo Max feature, the languages supported, and the no-penalty design intent.
• Shopify, "Sidekick" product page - Sidekick is included with Shopify plans with limits varying by plan, has access to store data, and takes action in the admin.
• Fin pricing page (read 20 July 2026) - $0.99 per resolution with a 50-outcome monthly minimum, one outcome per conversation, $9.99 per qualification.
• McCarthy Tétrault, "Moffatt v. Air Canada: A Misrepresentation by an AI Chatbot" - The February 2024 BC Civil Resolution Tribunal decision holding Air Canada liable for its chatbot's bereavement fare statement.
• ITV News, "DPD disables AI chatbot after it swears at customer" (19 January 2024) - DPD disabling the AI element of its chatbot after a customer induced it to swear and criticise the company.
• The Autopian, "Chevy Dealer's AI Chatbot Allegedly Sold A New Tahoe For $1" - The December 2023 prompt injection against a Chevrolet dealership chatbot.
• The Register, "Cursor AI support bot hallucinated its own company policy" (18 April 2025) - The invented one-device policy, the resulting cancellations, and the company's subsequent labelling of AI replies.
• The Markup, "NYC's AI Chatbot Tells Businesses to Break the Law" (29 March 2024) - The MyCity chatbot giving business owners advice that would have breached housing and labour law.
• STAT, "Woebot Health shuts down pioneering therapy chatbot" (2 July 2025) - Retirement of the consumer Woebot app on 30 June 2025 and the regulatory reasons given by its founder.
• TechCrunch, "The fall of Babylon" (31 August 2023) - Babylon Health's US Chapter 7 filing and wind-down, the reason it was removed from this list.
• CNBC, "McDonald's to end AI drive-thru test with IBM" (17 June 2024) - The end of the McDonald's and IBM automated order taking test, switched off in all restaurants by 26 July 2024.
• CBC News, "Kik messenger service closing down, cutting 90 staff" (2019) - Kik's announced shutdown, the reason the H&M Kik bot cannot be used and should not appear on current lists.
• Anthropic engineering, "Building effective agents" (19 December 2024) - The definition separating agents, which direct their own processes and tool use, from workflows running predefined code paths.