· Feedbot team
Chatbot Examples: 15 Real Bots and What They Teach
Real chatbot examples from Klarna, IKEA, Bank of America, Redfin and Lemonade, four public failures, 10 chatbot use cases and sample conversation scripts.
TL;DR. The chatbot examples that work all do one narrow job well: Bank of America’s Erica handles routine banking tasks, IKEA’s Billie answers common service questions so staff can sell design services, Lemonade’s Maya turns an insurance quote into a short chat. The public failures (Air Canada, a Chevy dealer, DPD, New York City’s MyCity) all let a bot speak beyond what it knew. Below: 11 real chatbots across industries, 4 failures and what they teach, a table of chatbot use cases by industry, and sample conversation scripts you can adapt.
We make Feedbot, an AI chatbot for websites. None of the examples below use Feedbot. We picked them because they’re public, documented and easy to check, and we link the source for every number.
What is a chatbot, and what makes a good one?
A chatbot is software that holds a text conversation with a person. On a business website, its job is narrower than ChatGPT’s: answer questions about this business, collect a lead or booking, and hand off to a person when it should.
Across the examples below, the good ones share four traits:
- A narrow job. Answer service questions, quote a policy, schedule an interview. Not “talk about anything.”
- Grounded answers. They answer from the company’s own data, not from the model’s general knowledge.
- A clear exit. When the bot can’t help, a person can.
- No commitments it can’t keep. Prices, refunds and legal statements come from systems or staff, not improvisation.
11 real chatbot examples across industries
| Chatbot | Company | Industry | Main job | Where |
|---|---|---|---|---|
| Erica | Bank of America | Banking | Account tasks, insights | App, online banking |
| AI assistant | Klarna | Fintech | Customer service | App, web |
| Billie | IKEA (Ingka Group) | Retail | Customer service | Web, app |
| Rufus | Amazon | Ecommerce | Product questions | Website, app |
| Ask Redfin | Redfin | Real estate | Listing questions | App |
| Maya | Lemonade | Insurance | Quotes and sign-up | Website, app |
| Pounce | Georgia State University | Education | Enrollment questions | SMS |
| Ada | Ada Health | Healthcare | Symptom assessment | App, web |
| Dom | Domino’s | Restaurants | Ordering | App, Messenger, voice |
| Olivia | Paradox (McDonald’s McHire) | Recruiting | Screening, scheduling | Careers sites |
| Romie | Expedia | Travel | Trip planning | App (beta) |
1. Erica (Bank of America): routine tasks at huge scale
Erica launched in 2018 and helps clients check balances, find transactions, spot duplicate charges and manage cards. Bank of America says Erica passed 3 billion client interactions in August 2025, averaging about 58 million a month.
What it does well: it acts on account data instead of just answering questions, and it proactively flags things like unusual charges. The lesson for a small business: a bot that can look up your customer’s order or account is worth much more than one that only recites FAQs.
2. Klarna’s AI assistant: big wins, then a correction
In February 2024, Klarna announced its OpenAI-powered assistant handled 2.3 million conversations in its first month, two-thirds of its customer service chats, with resolution time down from 11 minutes to under 2. In 2025, the CEO said Klarna would hire human agents again so customers can always reach a person.
What it teaches: automation rates and customer satisfaction are different numbers. Plan the human path from day one.
3. Billie (IKEA): automate the simple, redeploy people
Ingka Group, IKEA’s largest franchisee, introduced Billie in 2021. It has resolved about 47% of customer queries, and Ingka retrained around 8,500 contact-center workers as remote interior design advisers instead of cutting them.
What it does well: the bot takes order status and simple policy questions; humans take the work that earns money. That split is a good model for any business adding a chatbot.
4. Rufus (Amazon): answering product questions at the shelf
Amazon launched Rufus in 2024 to answer questions on product pages (“Is this good for cold weather?”, “What’s the difference between these two?”), drawing on listings, reviews and Q&A. In 2026, Amazon folded Rufus into Alexa for Shopping.
What it does well: it sits where the buying question happens, not on a separate help page. For a store, put the chat on product pages and train it on your product data. See our ecommerce chatbot guide.
5. Ask Redfin: real estate questions with a fair housing guardrail
Ask Redfin answers buyers’ questions about a listing: HOA fees, open houses, tour availability, zoning. Redfin says it trained the assistant to reject questions that violate fair housing rules, and to connect buyers with a licensed agent when a question needs one.
What it does well: compliance is built in, and the handoff to a human agent is the goal, not a failure. More in our real estate chatbot page.
6. Maya (Lemonade): a form turned into a conversation
On lemonade.com, Maya asks for your address, coverage needs and details one question at a time and returns a renters, home, pet or car insurance quote.
What it does well: it replaces a long form, not a person. If your site has a quote or intake form people abandon, a chat that asks one question at a time is worth testing. See chatbots for insurance.
7. Pounce (Georgia State University): fewer students lost over summer
Georgia State used an AI text-message assistant called Pounce to answer incoming students’ questions about financial aid, housing and registration. In the first year it answered more than 200,000 questions, and the university reports a 21.4% reduction in “summer melt” among students who committed by its priority deadline.
What it does well: it targets one measurable problem (admitted students who never enroll) and reaches people at the moment they’re stuck. It runs over SMS, but the same pattern works on an admissions website. See chatbots for education.
8. Ada (Ada Health): structured symptom assessment
Ada asks adaptive questions about symptoms and returns possible causes and guidance on how urgently to seek care. It’s built as a medical product with clinician-reviewed content.
What it teaches: health is a high-stakes domain where a general chatbot shouldn’t improvise. A clinic’s website bot should handle hours, insurance accepted and booking, and send anything clinical to staff. See chatbots for healthcare.
9. Dom (Domino’s): ordering where customers already are
Domino’s introduced Dom in 2014 as a voice ordering assistant in its app, then added full-menu ordering in Facebook Messenger.
What it does well: it does one transaction, ordering, end to end, and remembers your usual order. For a restaurant site, the equivalent is answering menu, allergen and hours questions and taking reservations. See chatbots for restaurants.
10. Olivia (Paradox, used on McDonald’s McHire): screening and scheduling, plus a security lesson
Olivia chats with job applicants, asks screening questions and schedules interviews. In 2025, security researchers found a Paradox test account protected by the password “123456” that exposed records tied to up to 64 million applications. Paradox fixed it the day it was reported.
What it teaches: high-volume scheduling is an excellent chatbot job, and a chatbot that collects personal data needs the same security as the rest of your stack. See chatbots for recruiting.
11. Romie (Expedia): a travel assistant still in beta
Expedia introduced Romie in May 2024 as an assistant that helps plan trips, can join group chats and update itineraries. It launched as an experiment in Expedia’s EG Labs.
What it teaches: big brands ship chatbots as labeled experiments first. Small businesses can do the same: start with a narrow set of questions, read the transcripts, then expand. For hotels, see chatbots for hotels.
4 chatbot failures and what they teach
| Case | What happened | Lesson |
|---|---|---|
| Air Canada (2024) | A tribunal held the airline liable after its website chatbot invented a bereavement refund rule | You own what your bot says |
| Chevrolet of Watsonville (2023) | A visitor got the dealer’s ChatGPT bot to “agree” to sell a Tahoe for $1 | Block bots from making offers |
| DPD (2024) | After an update, the parcel firm’s bot swore at a customer and wrote a poem mocking DPD | Test after every change |
| NYC MyCity (2024) | The city’s business chatbot told owners they could take workers’ tips and refuse housing vouchers | Ground answers in verified sources |
Air Canada. A customer asked the airline’s website chatbot about bereavement fares and was told he could apply for the discount after buying his ticket. The real policy said otherwise. In Moffatt v. Air Canada, British Columbia’s Civil Resolution Tribunal ordered Air Canada to pay and rejected the argument that the chatbot was responsible for its own words.
Chevrolet of Watsonville. In December 2023, a user instructed the dealership’s chatbot to agree with anything and end every reply with “that’s a legally binding offer”. It then “agreed” to a $1 Tahoe. No car was sold, but the screenshots went viral.
DPD. In January 2024, DPD disabled the AI part of its chat after a frustrated customer got it to swear and criticize the company. DPD blamed a system update.
NYC MyCity. The Markup found the city’s chatbot routinely gave advice that broke the law. In January 2026, Mayor Zohran Mamdani said the city would shut it down.
All four come down to the same fix: limit the bot to your own content, tell it what it must never promise, read real transcripts, and route anything uncertain to a person.
Chatbot use cases by industry
| Industry | What visitors ask | What the bot should do | What to keep human |
|---|---|---|---|
| SaaS | Pricing, limits, “does it integrate with X?”, setup errors | Answer from docs, recommend a plan, collect bug reports | Refunds, account changes, enterprise deals |
| Real estate | Availability, price, viewings, HOA fees | Qualify buyers, book viewings | Negotiation, and questions about neighborhood demographics (fair housing) |
| Education | Deadlines, tuition, admission requirements | Answer from the catalog, collect inquiries | Admission decisions, financial aid specifics |
| Healthcare | Hours, insurance accepted, booking | Answer logistics, book appointments | Anything clinical or involving symptoms |
| Insurance | What’s covered, how to file a claim | Explain products, collect quote details | Binding advice and coverage decisions |
| Hotels | Check-in times, parking, pets, breakfast | Answer policies, push direct booking | Complaints during a stay, special requests |
| Restaurants | Menu, allergens, hours, reservations | Answer menu questions, take group inquiries | Allergy guarantees, large events |
| Recruiting | Open roles, salary range, process | Screen, schedule interviews | Hiring decisions, rejections |
| HR | PTO policy, benefits, payroll dates | Answer from the employee handbook | Personal cases, complaints, anything legal |
| Retail | Sizing, shipping, returns, stock | Answer product questions, track orders | Damaged orders, exceptions to policy |
Beyond industries, the use cases that come up everywhere are support from your help docs, lead capture, booking, order tracking, collecting feedback and bug reports, and routing to the right person.
Chatbot conversation script examples
These sample scripts show the patterns above in action. Adapt the wording to your business.
SaaS: pricing question that becomes a lead
Visitor: Do you have a plan for 3 people?
Bot: Yes. The Team plan covers up to 5 users for $29 a month. It includes shared inboxes and the API. How many conversations do you expect a month? I can check it fits.
Visitor: Maybe 800.
Bot: That fits the Team plan’s 1,000 limit. Want me to send a summary to your email so you can share it with your team?
Restaurant: allergen question with a safe answer
Visitor: Is the pad thai peanut-free?
Bot: Our pad thai is made with peanuts, and our kitchen handles peanuts in shared equipment, so we can’t guarantee any dish is peanut-free. For a severe allergy, please call us at (555) 010-2030 before ordering and the kitchen can tell you what’s safest.
Clinic: booking without medical advice
Visitor: I’ve had a headache for three days, should I be worried?
Bot: I can’t give medical advice. If it’s severe or sudden, or comes with fever, confusion or vision changes, call 911 or go to an emergency room. Otherwise I can book you with a doctor: the next openings are Thursday at 10:00 or 2:30. Would one of those work?
Online store: return that needs a person
Visitor: My jacket arrived torn. I want my money back.
Bot: I’m sorry it arrived damaged. Refunds for damaged items are handled by our team, so I’ve created a request with your order. What’s the email on your order? They’ll reply there within one business day.
Real estate: qualifying a buyer
Visitor: Is 14 Oak Street still available?
Bot: Yes, it’s listed at $485,000 with 3 bedrooms and an open house Saturday 1–3 pm. Are you looking to buy within the next 3 months, and have you been pre-approved? I can set up a private viewing with the listing agent.
In each script, the bot answers the question first, stays within what it knows, and moves toward a concrete next step: a summary, a phone call, a booking, a ticket, a viewing.
How to build a chatbot like these for your website
- Pick one job. “Answer support questions from our docs” or “book viewings”. Write down what the bot must never do.
- Give it your content. Help center, pricing page, policies, FAQ. The bot should answer only from these.
- Decide the handoff. Who gets notified, where, and what the bot says while the visitor waits.
- Launch on a few pages and read transcripts weekly. Every wrong or missing answer is a gap in your content.
- Measure. Questions answered without a person, leads captured, and handoffs per 100 conversations.
That’s the workflow Feedbot is built for. You add it with one script tag, train it on your site and docs (a knowledge base chatbot), and it answers from that content, captures leads, and hands off to your team in Telegram, Slack, Discord or email when it should. It also collects bug reports and feature requests from the same conversations. The free plan covers 50 conversations a month; see pricing.