- AI chatbots work well for predictable FAQs and for sorting inquiries, but they fail in complex or emotionally charged situations.
- A badly designed chatbot is worse than no chatbot at all: it creates frustration and damages your brand.
- Success depends more on conversation design and managing expectations than on how sophisticated the AI model is.
A few months ago, a client proudly showed me the AI chatbot he had just installed on his website. “Now users can ask anything and get instant answers”. I asked whether he had reviewed the conversations. We opened the history. One customer had written: “My order hasn’t arrived and I need it for tomorrow”. The chatbot replied with generic information about standard delivery times. The customer tried 3 more times, more frustrated each time, until they gave up. They never got to talk to a person. Lost order, angry customer, dented reputation.
AI chatbots are a double-edged sword. Used well, they reduce the support workload and improve the experience. Used badly, they turn small problems into customer relationship crises.
When AI chatbots help
Chatbots shine in specific cases:
Predictable, factual FAQs. Opening hours, returns policy, shipping costs, how to reset a password. Documented information that doesn’t change with context.
Initial triage and routing. Identifying the type of inquiry and sending it to the right department or support level: “Is this about sales, technical support or billing?”. That means fewer transfers.
24/7 availability for simple inquiries. When the user just needs a quick piece of information outside office hours.
Collecting information before contact. Asking for the order number, account email and a description of the problem before handing over to a person. The agent starts with context.
Step-by-step guided processes. “To set up X, do this, then this, then this”. Clear, ordered instructions.
When AI chatbots drive customers away
These are the scenarios where we see chatbots do more harm than good:
Emotionally charged situations. Angry, stressed or frustrated customers. A chatbot that doesn’t recognize the emotion and gives a generic reply makes things worse. When someone writes “this is unacceptable, I’ve been waiting for an hour”, they don’t want an FAQ answer: they want empathy and a solution.
Complex problems with context. “My order from 3 weeks ago arrived damaged, the replacement is wrong too and now I need it for an event tomorrow”. Situations like this require understanding the history, prioritizing and making decisions outside the standard procedure.
Company mistakes that call for an apology. When the company has got something wrong, customers want the mistake acknowledged and put right. A chatbot that “understands your frustration” but can’t do anything tangible breeds cynicism.
High-value decisions. Inquiries about orders worth thousands of euros, contracts or legal matters. Customers want the certainty of a person, not a probabilistic answer.
Hidden exits. When it is hard or impossible to reach a person. The “talk to an agent” option only appears after 5 failed attempts to get the bot to solve it. This makes people furious.
Conversation design: the key to success
The underlying AI model matters less than you think. What matters is conversation design:
Clear expectations from the start. Users should know from the first message that they are talking to a bot and what it can do: “I’m the virtual assistant. I can help with orders, returns and your account. For anything else, I’ll put you through to a person”. On top of that, the EU AI Act already includes the obligation to inform people that they are interacting with an AI system.
An easy way out to a person. Always visible, always working. “Talk to a person” one click or one message away, not buried in menus.
Context preserved in the handoff. When the conversation passes to a person, the agent sees the full history with the bot. The customer should never have to repeat themselves.
Explicit limits. “I can’t change orders that have already shipped, but I’ll put you through to someone who can”. The bot shouldn’t try to answer everything.
The right tone. More formal or more casual depending on the brand and context, but never fake enthusiasm. A “Great! I love helping you!” when the customer is complaining puts people off.
Technology versus strategy
People selling AI chatbots focus on technical capabilities: “understands context”, “learns from conversations”, “runs on the latest AI model”. But success depends on your customer service strategy:
Which cases the bot covers and which it passes to a person. This is a business decision, not a technical one, and it requires understanding what your customers value.
Escalation rules. If the bot doesn’t solve it within 2 exchanges, the conversation is handed over automatically. If no agent is available, decide what message is shown and what alternative is offered (a call back within a set time, for example).
Improvement loop. How failed conversations are used to improve the bot or the underlying documentation.
How success is measured. Not by the “number of conversations resolved by the bot”, but by comparing customer satisfaction with the bot and with people. If the bot resolves 80% of inquiries but only satisfies 40% of the people who use it, it’s a failure.
Real costs versus promised costs
It is common to come across promises like “cut your support costs by 70%”. The reality is more nuanced:
Direct costs: the chatbot license, usage of the AI model’s API (charged by tokens, in other words by the volume of text processed) and integrations with your systems.
Indirect costs: time spent on conversation design, maintaining the knowledge base, reviewing failed conversations and handling complaints from frustrated customers.
Risks: customers lost through a bad experience, brand damage and missed sales because the bot didn’t spot buying intent. There is legal risk too: in 2024, a Canadian tribunal ordered Air Canada to compensate a passenger after the airline’s chatbot gave him wrong information about its refund policy. What your bot says, your company says.
A rough calculation: if the bot handles 30% of inquiries well and routes the rest correctly, you probably save money. If it resolves 60% but leaves 1 in 5 of the customers who use it frustrated, you can lose money. We go into more detail in automating customer service with AI: the real cost versus the promised one.
Alternatives to an AI chatbot
Before rolling out an AI chatbot, consider:
A better search. For information queries, a semantic search (one that understands synonyms and questions phrased in different ways) is sometimes enough, and it frustrates people less than a chatbot.
A rule-based chatbot. If your inquiries are predictable and limited, a decision-tree chatbot (no AI) is more predictable and cheaper.
Smart forms. A “Tell us what’s happening” form with structured fields that guide the user and route the inquiry automatically. For clear-cut cases, it’s faster than a chat. If your current form is already failing, start there: why your contact form isn’t generating leads.
People only, but more efficient. Sometimes investing in better tools for your team (saved replies, templates, quick access to customer information) has a better return than automating.
Frequently asked questions
Should I replace my support team with an AI chatbot?
No. Chatbots complement people, they don’t replace them. The model that works is a bot for triage and FAQs, and people for complex cases. You’ll reduce the volume of simple inquiries, but you’ll still need your team.
How long does it take to “train” an AI chatbot?
It isn’t training in the traditional sense. There’s an initial setup (a few weeks) and ongoing tuning (every month): reviewing conversations, spotting failures and updating the knowledge base. It’s ongoing work, not a one-off project.
What happens when the chatbot says something wrong?
It depends on how you configure it. Some systems include a notice along the lines of “I’m an AI assistant, please check important information”. For critical inquiries (prices, terms, policies), the sensible approach is for the bot to always pass the query to a person rather than risk a wrong answer.
Is an AI chatbot better than a rule-based one?
For limited, predictable cases, rules are cheaper and more predictable. For broad cases, where users can ask “anything”, AI has the edge. But most companies don’t need to answer “anything”: they need to answer their 50 most frequent questions well.
Conclusion
AI chatbots aren’t inherently good or bad: they are tools, and the outcome depends on how they are used.
Success doesn’t come from how sophisticated the model is, but from smart conversation design, good expectation management and the humility to recognize when a person is needed.
If you’re thinking about rolling out an AI chatbot and want help designing the strategy (what the bot covers, when it hands over to a person, how success is measured), get in touch. I’d rather help you build a modest bot that works well than an ambitious one that frustrates your customers.