You've probably heard the stories. A chatbot told a customer they could return a product under a policy that didn't exist. Another invented a discount. One airline's bot promised a bereavement fare that wasn't real — and the company was held to it in court. These incidents are real, and the skepticism they've created is reasonable.
At the same time, chatbots are handling millions of customer interactions a day without incident. The difference isn't luck. It's in how they're built and what they're asked to do.
What Makes a Chatbot Actually Work
The chatbots that cause problems share a common pattern: they're given too much freedom and too little grounding. They're pointed at a general AI and told to "be helpful" — so they are, including when that means making things up.
A chatbot that works reliably is a narrow one. It knows your business specifically — your products, your policies, your prices, your opening hours — and nothing else. When a customer asks something outside that scope, it says so and offers to connect them with a person. It doesn't guess. The moment a chatbot starts reaching beyond what it's been explicitly trained on, you've lost control of what it says.
Good chatbots are also transparent. Customers don't mind automated responses — they mind slow, unhelpful, or misleading ones. A chatbot that answers quickly, accurately, and hands off cleanly when needed tends to get better feedback than a human who is overloaded and distracted.
How Much Can Realistically Be Automated
This depends almost entirely on what kind of business you run and how your customers interact with you.
A business with a high volume of repetitive, predictable enquiries — the same questions asked dozens of times a day — is a good candidate. Online retail, service businesses with standard packages, clinics handling appointment bookings, landlords managing rental enquiries: these tend to see real benefit. A large proportion of incoming contacts follow the same few patterns, and those patterns can be handled reliably by software.
A business built on relationships, bespoke work, or complex sales conversations is a different story. If most of your customer interactions require context, negotiation, or judgement — a consulting firm, a bespoke manufacturer, a financial adviser — a chatbot covering the first line of contact might handle a small fraction of enquiries usefully, but the bulk of the value comes from human conversations. Automating more than the basics there can feel impersonal in a way that damages the relationship rather than supporting it.
The honest answer to "how much can we automate" is: map your actual incoming contacts, categorise them, and see how many fall into simple, repeatable patterns. If the answer is most of them, a chatbot is worth considering. If it's a minority, the return on investment shrinks accordingly.
When a Chatbot Is Worth Building — and When It Isn't
A chatbot is probably worth building if your team regularly spends time answering the same questions, if customers contact you outside business hours and leave without getting a response, or if you're missing enquiries because you can't get to them fast enough.
A chatbot is probably not worth building if your business runs on relationships and personal service that customers value specifically because it's human, if your products or services are complex enough that most enquiries need real discussion, or if you don't have clear, well-documented information to train it on. A chatbot trained on vague or inconsistent information will give vague or inconsistent answers — and that reflects on the business, not just the software.
There's also a maintenance question. Chatbots are not set-and-forget. When your prices change, your policies change, or you add a service, the chatbot needs to be updated. If nobody is going to own that, the chatbot will gradually drift out of sync with reality — and you'll end up with exactly the kind of misinformation problem you were trying to avoid.
Let's Talk
If you're weighing up whether a chatbot makes sense for your business, we're happy to think it through with you — including whether the answer is no. A straightforward conversation about your actual enquiry patterns usually makes it clear pretty quickly which way it goes.