01What a badly built bot does at the moment it fails
Every chatbot meets a question it cannot answer. The gap is not the failure. What the bot does next is. The first bad habit is guessing. In 2024 a British Columbia tribunal ruled on a case where Air Canada's website chatbot told a passenger he could apply for a reduced bereavement fare within 90 days of the ticket being issued. Air Canada argued it could not be held liable for what the chatbot said. The tribunal rejected that, because the chatbot was still part of the airline's website, and ordered Air Canada to pay damages, fees and interest. The second habit is the loop. The US Consumer Financial Protection Bureau's 2023 report on chatbots in consumer finance describes customers led into continuous loops of repetitive, unhelpful replies with no offramp to a human. The third is quieter: the person exists, but the route to them is buried or slow. In each case the customer learns one thing, that the bot is in the way.
- Guessing: a confident wrong answer is still your answer
- Looping: the same apology three times, with no way out
- Hiding the human: the person exists, but the route to them is buried
- Staying silent: nobody in your business learns the question was asked
02What a well-built bot does instead
A well-built bot treats not knowing as a normal event with a script of its own. It says plainly that it cannot answer that one. It tells the customer what happens next and when, because 'someone will be in touch' with no time attached reads as a brush-off. It passes the conversation to a person with everything already said, so the customer does not repeat themselves. It keeps the question for you, since each unanswered question is either a gap in the knowledge base or a case that should always go to a person. Outside working hours the honest answer is a time, not an invented reply. Some topics should skip the bot's attempt entirely: a complaint, a refund, anything with legal or safety risk. Our AI assistant page describes it the same way. The assistant says it does not know and routes the conversation to a person, and you decide what triggers the handoff, such as the type of topic or the time of day. How those rules are set is covered in our post on one knowledge base and one handoff.
- Says plainly that it does not know that one
- Names the next step and a time
- Hands over the whole conversation, so nobody starts again
- Sends complaints, refunds and legal or safety topics straight to a person
- Shows the question in a weekly review, so it becomes a new answer or a new rule
03Questions to put to a vendor
Do not ask a vendor whether the bot hands off to a person. Every vendor says yes. Ask to see it happen. Give them three questions their bot should not be able to answer: one from outside your business, one about a price that does not exist, and one that is an angry complaint. Then watch what it does. After that, ask the plain questions below. A vendor who built this properly has the answers ready and shows you a list of real conversations, not a slide. Our post on what to ask before you buy an AI chatbot covers price, setup and training data. These questions cover only the moment the bot fails.
- Show me the bot failing: what does it say to a question outside its knowledge?
- What triggers a handoff, and can I change the triggers or must I ask you?
- Who receives the conversation, where, and do they see the whole thing?
- What does the customer see outside working hours?
- How do I see the questions the bot could not answer?
- Can the customer ask for a person at any point?
04Where PROOF48 is the wrong choice
PROOF48's AUTOMATE plan costs 2,997 EUR a month and includes an AI website chatbot that qualifies and books, AI DM and email auto-responders, booking automation with a CRM and two custom automations a month. It does not cover WhatsApp or phone. The handoff is part of the build: we test edge cases and unhappy clients before launch, then you review conversations weekly, sorted into answered, escalated and unanswered. We publish no figure for how often our chatbot hands off, because we have not measured it for a client and will not invent one. We are the wrong choice if every conversation in your business needs a human judgment call. The bot would hand off constantly and you would have bought a form with extra steps. Our post on whether an AI chatbot is worth it draws that line in detail. And if your questions arrive mostly on WhatsApp or by phone, AUTOMATE is not the plan. A WhatsApp assistant is a separate build, priced on request.
- AUTOMATE: 2,997 EUR a month, website chatbot, DM and email auto-responders, booking with a CRM
- Handoff is tested with edge cases and unhappy clients before launch
- WhatsApp and phone sit outside AUTOMATE
- Wrong fit: every conversation needs a human judgment call
- No handoff-rate figure published, because none is measured
A visible route to a person on a hotel site
Santorini Hotel's site in Eforie Nord, built by PROOF48, publishes separate reception, restaurant and complaints phone lines on the site and in its structured data, next to a booking enquiry flow. It is not a chatbot and we claim no chatbot there. The idea carries over anyway: a guest with a complaint should not have to guess whom to call, and a bot should hand over in the same spirit, to the right person for that kind of question. We publish no booking or occupancy figures for the hotel.
See the build: Santorini HotelQuestions, answered
Should a chatbot ever say it does not know?
Yes. 'I cannot answer that one, here is what happens next' beats a confident guess. In the Air Canada case the guess became the company's problem: the tribunal ordered the airline to pay.
How often should a bot hand off to a person?
We give no figure, because we have not measured it for any client. Our AI assistant page says the first month is for learning and escalations drop after that, but how fast depends on how stable your questions are.
Does a handoff mean the customer has to phone someone?
No. It means a person takes over the conversation. Which channel they continue on is something to settle before launch. AUTOMATE covers website chat, DM and email, not WhatsApp or phone.
Where PROOF48 fits
When a chatbot cannot answer, the test is what it does next. A good one admits it, names the next step and a time, and hands the whole conversation to a person. A bad one guesses, loops or hides the human. Ask any vendor to show the failure, not the demo. PROOF48's AUTOMATE plan, 2,997 EUR a month, includes a website chatbot with handoff rules you choose, tested before launch. It does not cover WhatsApp or phone.
Cosmin Tiu · Founder, PROOF48
Cosmin has worked in online marketing and software development since 2018, and has built with AI every day since January 2023. He still writes code from scratch daily, in Python, React, Rust and C#. PROOF48 ships business websites in 48 hours and invoices afterwards, so every opinion on this blog was paid for in real builds, not in theory.