“We already tried a chatbot and it was a disaster.”
It’s the objection that comes up most when we explain what we do, and the interesting part is that it’s usually true. Plenty of businesses tried a button-based chatbot a few years ago, customers got frustrated, somebody eventually switched it off, and a fairly reasonable conclusion stuck around: automating customer service drives people away.
The problem is that the conclusion was drawn from a different technology. An AI agent and a rules-based chatbot are alike in that they both reply to messages, and that’s where the resemblance ends. It’s worth understanding the difference before writing off the whole category, if only to know what you’re turning down.
Same question, two answers
A customer writes “do you have anything tomorrow afternoon?”. A rules-based chatbot replies “press 1 to book, 2 for pricing, 3 for our address”. An agent replies “yes, I have 4:00 and 5:30 pm open — which works better?”.
The first one didn’t understand the question and returned a menu. The second understood what the customer wanted, checked the actual calendar, and offered real options, which is what the customer expected in the first place because it’s what would have happened if a person had answered.
Why chatbots fail that way
A rules-based chatbot is a decision tree somebody designed in advance: if the user says this, reply with that; if they press 1, show this menu. There’s nothing else under the hood.
Every one of its limitations follows from that, and they’re fairly predictable. It only covers what someone anticipated, so any question outside the tree ends in “sorry, I didn’t understand.” It depends on keywords, so if you programmed “hours” but the customer types “what time do you open,” it can miss. It doesn’t look anything up, so it answers with fixed text without knowing whether there’s actually availability. And it doesn’t execute: it can tell you to call to book, but it can’t book.
None of this means the technology is bad. For a three-option menu that never changes, a rules-based chatbot works fine and costs very little. The trouble starts when it’s used to handle real customers, who don’t write the way the tree expected.
What an agent does differently
An agent uses a language model to interpret what the customer wants, and it has access to tools to do something about it. In a booking conversation, four things happen, all in under two seconds.
First it works out the intent: it isn’t matching words but understanding that “any room Thursday?”, “I want an appointment,” and “when can you see me” are the same request, typos and missing punctuation included. Then it queries your actual systems — your calendar, your inventory, your CRM — and reads the current state, which is why it can say “I have 4:00 and 5:30” instead of asking you to call. Then it answers from what it found rather than what it imagines. And finally it executes: blocks the slot, sends the confirmation, schedules the reminder. The conversation ends with a booked appointment rather than a promise that someone will call back.
There’s a fifth behaviour that matters as much as the other four and gets mentioned least in demos: when something falls outside its scope, it hands off to a person along with the conversation context. An agent that improvises when it doesn’t know is worse than no agent at all, because it hands your customers false information in the confident tone of something that does know, and the customer has no way to tell the difference.
What an agent shouldn’t do
Worth saying just as plainly, because this category gets sold with a lot of exaggeration. An agent doesn’t replace professional judgement: it can book a dental appointment but not decide on a treatment, qualify a real estate lead but not negotiate the price of a property. It also shouldn’t make irreversible decisions on its own, because charging, cancelling services, or changing sensitive records needs human confirmation or at minimum an audit trail somebody can review afterwards.
And most importantly, it doesn’t fix a broken process. If your calendar is a mess or nobody’s clear on who handles what, automating first contact will produce disorder faster. Automation amplifies the process you already have, in the direction it was already heading.
Which one you actually need
A rules-based chatbot is enough if your enquiries are few and always the same, you don’t need to check real-time availability, and the goal is delivering fixed information like hours and address. It’s a perfectly reasonable tool for that, and buying something more expensive would be wasting money.
An agent makes sense when you get high message volume, your customers write in natural and varied language, you need to check a calendar or inventory, you want conversations to end in a concrete action, and a meaningful share of messages arrives outside working hours. And if your volume is low — say under twenty enquiries a month — neither one is the answer, because that isn’t a capacity problem but a demand problem, and the money is better spent elsewhere.
Three questions before you sign anything
If you’re evaluating vendors, three questions separate the serious ones from the rest quickly.
What happens when the agent doesn’t know the answer? Ask to see it live rather than be told about it, because how an agent handles what it doesn’t know determines whether it helps you or costs you customers. What systems does it connect to, and how? If the answer is vague, it probably doesn’t connect to anything and you’re buying a chatbot with better copywriting. Who owns the conversations and the data? Your customers’ information is yours, and if that isn’t clear in the first meeting, that’s a signal.
In the end the distinction between chatbot and agent isn’t a technical one, and that’s exactly why it matters: if your previous experience was with a button menu, the conclusion you drew about automating customer service applies to something different from what’s available now. You might land on the same conclusion after a second look, but at least it’ll be about the right technology.
At Katalis we build agents connected to WhatsApp and your calendar. If you’d like to see one running against your own business cases, get in touch.