If you run a business with a website, you have probably been offered a chatbot. The pitch sounds good: customers get instant answers, you save on staff time, and the AI handles the routine questions so your team can focus on the complicated ones. But a new research paper raises a question that most chatbot vendors would rather you did not ask: how do you actually know if your chatbot is any good?
The paper, titled “Operationalising Multi-Dimensional Evaluation for Conversational Agents,” comes from researchers who built and tested a large-scale evaluation system for retail chatbots. They processed more than two million customer interactions and found that evaluating a chatbot properly requires looking at far more than whether it answered the question. You also need to check whether it was helpful, truthful, clear, and appropriately toned for the customer. These are different things, and a chatbot can ace one while failing another.
The Problem With “It Answered the Question”
The simplest way to test a chatbot is to see if its answer matches the right keywords. The researchers call this “lexical-overlap metrics,” and it is the lowest bar. A chatbot can match keywords while giving terrible advice. Worse, it can sound confident while being wrong, and most customers will not challenge it because it looks like an official company response.
The paper’s solution is a structured evaluation pipeline that scores every chatbot interaction across multiple dimensions: helpfulness (did the customer get what they needed?), truthfulness (was the information accurate?), clarity (was the answer easy to understand?), tone alignment (was the language appropriate for the situation?), and intent recognition (did the chatbot understand what the customer actually wanted?).
This is more work than running a simple keyword test. But the researchers showed it works at scale, processing roughly 50,000 interactions per day across 14 different customer intents, 18 business domains, and 156 sub-intents. Their system achieved a 93 percent accuracy rate when compared against human evaluators.
What This Means for a Small Business Owner
If you are considering adding a chatbot to your letting agency, trade business, or retail site, the lesson is simple. Do not accept a demo where the vendor shows you a few neat examples. Ask how they evaluate performance across real customer interactions. Ask whether they track truthfulness separately from how often the bot answered. And ask what happens when the chatbot does not know the answer — does it admit it, or does it make something up?
The paper found that “hellpfulness” is not the same as “accuracy.” A chatbot can be very helpful while being wrong, and very accurate while being unhelpful. If your vendor is only measuring one of these, they are giving you an incomplete picture.
Practical Advice Before You Buy
When trialling a chatbot for your business, do these three things. First, run your own test scenarios using actual customer questions from your email or phone records, not the examples the vendor provides. Second, ask a real customer or two to use it and tell you how it felt, not just whether it worked. Third, check what happens when the chatbot hits something it does not understand. A good chatbot escalates to a human gracefully. A bad one tries to bluff.
The research confirms what anyone who has used a bad chatbot already knows: the technology is impressive, but it needs proper oversight. The businesses that get value from chatbots are the ones that test them properly, not the ones that install them and forget.