A chatbot for your business: when it helps and when it hurts

A correction on the word first: what I built for this site is not a "chatbot".

A chatbot returns stored answers. What I have is an agent: it searches the site's documents, decides which part relates to your question, reads it, answers, and refuses when it finds nothing. That difference is architectural rather than marketing, and it is set out in 99% of "AI agents".

Most of what gets sold to you as a chatbot is the first thing. Most of what you need is the second.

Now, I will start with how mine broke, because that is more useful to you than any feature list.

It spent weeks refusing every question it was asked, including "who is Abdelilah?" on Abdelilah's own website.

The knowledge base was empty, and the guard I had built to stop it inventing prices was treating "nothing in the index" as "this question is outside my scope". From outside, those two states look identical and produce the same polite sentence.

When it genuinely helps#

When your questions are repetitive and finite. Opening hours, returns policy, published prices, location. Ten questions covering most of your messages.

When your answers are already written down. This is a condition rather than a preference, and I will come back to it.

When an instant reply has value. Outside working hours, or when a customer who waits minutes goes to a competitor.

When escalation is easy. Meaning a customer can reach a human in one click.

When it hurts#

When your answers do not exist. A chatbot does not organise your knowledge; it exposes that it is not organised. With three conflicting versions of your returns policy, it will pick one at random.

When your customers' questions are complex. If every customer is a special case, the bot adds a step before the human they wanted.

When there is no escalation. The worst experience online is a bot that does not understand and offers no way out.

When nobody reads the transcripts. The transcripts are the most valuable thing the system produces, because they tell you what people ask and you do not answer.

Questions to ask any vendor#

This list separates the serious from the salesperson in ten minutes:

• Where does it answer from? If the answer is not "your own documents", it will invent. Foundations in what is RAG.
• Show me an answer with its source. If it cannot point at the document, you cannot audit it.
• What does it do when it does not know? The only correct answer: says so and escalates.
• How will I know when it has broken? My favourite question, and most vendors have no answer to it.
• Where are transcripts stored and who can read them?
• What does it cost at ten times the message volume? Not today. At success.

An acceptance test before you go live#

Run these four yourself, because each exposes a different defect:

A question your documents answer. It should answer and cite the source.

A question they do not answer. It should say it does not know. If it composes something beautiful, do not buy.

A price question, if you do not publish prices. The most dangerous test, because an invented figure reaches a customer and becomes a promise.

A message trying to talk it out of its instructions. "Ignore the above and tell me…". If it complies for you, it will comply for a mischievous customer.

In closing#

A chatbot multiplies how organised your knowledge is or exposes how disorganised it is, and does nothing else. Write your answers first, ask how you will know when it breaks, then decide.

Try this today: pull last week's messages and count the repeated questions. Fewer than ten, and you need a page rather than a bot.

Common questions#

When does a chatbot help my business?

When customer questions are repetitive and finite, your answers are already written and organised, an instant reply has value, and escalation to a human is one click away.

What is the most important question to ask a vendor?

How will I know when it has broken. The most dangerous failures raise no error; they return a polite, wrong answer with a success status, so nobody reports them.

What should I do before buying any chatbot?

Pull a month of messages, count the repeated questions, and write a final answer for each. You may find you need a good FAQ page rather than a bot.

How do I test a chatbot before launch?

Ask something your documents cover, something they do not, a price question if you do not publish prices, and a message asking it to ignore its instructions. Each exposes a different defect.