Nearly every business that asks us for a chatbot describes the same hope: it should handle the repetitive questions so the team can do real work. That is a reasonable goal and usually achievable. The trouble starts when the brief quietly expands into "and it should handle everything else too".
What they are genuinely good at
Answering the same forty questions. Opening hours, location, whether you deliver to a particular area, what is included in a package, how long something takes. These make up the bulk of incoming messages for most businesses, and a bot with access to your actual information answers them correctly and instantly, at any hour.
Qualifying before a human gets involved. Collecting the name, the requirement, the rough budget and the timeline, so that when your team picks the conversation up they are not starting from nothing. This is often more valuable than the answering itself.
Booking and scheduling. Connected to a real calendar, a bot handles appointment booking well because the task has clear rules and a definite end state.
Being awake. A large share of enquiries arrive in the evening and at weekends. An instant, accurate reply at 10pm frequently beats a better reply at 11am the next day, because by then the customer has messaged two competitors.
Where they fall over
Anything involving money it can change. Refunds, discounts, cancellations with a fee. Language models are agreeable by nature and a determined customer can talk one into offering things you never authorised. Keep those decisions behind a human.
Facts nobody wrote down. A model given no source will still produce an answer, and it will sound just as confident when it is wrong. If your policy is not written somewhere the bot can read, do not expect it to be represented correctly.
Upset people. Someone whose order did not arrive does not want an efficient assistant. Detect frustration and hand over fast.
Anything genuinely unusual. The edge cases are exactly the conversations where getting it wrong costs you the customer.
The measure of a good bot is not how many questions it answers. It is how quickly it recognises the ones it should not.
The part everybody skips
Most of the work in a chatbot project is not the model. It is writing down what your business actually does, accurately, in a form the bot can use. Prices, policies, exceptions, the things everyone in the office knows but nobody documented.
This is tedious and it is also the entire difference between a bot that helps and one that invents. Businesses that do this properly get something useful. Businesses that skip it get a confident liar wearing their logo.
A side benefit worth mentioning: the exercise of writing it all down usually surfaces two or three internal contradictions nobody had noticed. That is useful whether or not you ever launch the bot.
How to tell if you should build one
Before committing, spend a week doing this: keep a note of every enquiry that comes in, and mark each one as either repetitive or genuinely specific.
- If most are repetitive, a bot will help immediately
- If most are specific, a bot will frustrate people, and better contact routing is the real fix
- If volume is low enough that one person handles it comfortably, you probably do not need one at all
That week of counting is worth more than any vendor demonstration, including ours.
Start narrow
The bots that work start with one job. Answer FAQs, or book appointments, or qualify leads. Not all three. Get one right, measure how often it resolves without a handover, then widen the scope.
Bots that launch trying to do everything tend to do all of it at about sixty percent, which is worse than useless because it teaches customers not to trust it.