Best Practices· 6 min read

Avoiding the Robotic Feel Without Sacrificing Accuracy

How to get natural chatbot replies with short answers, plain words and good follow-up questions, while keeping every fact tied to your own content.


A visitor asks whether you work on Saturdays. The assistant replies with a paragraph that opens "Thank you for your enquiry regarding our operating hours", lists every day of the week, repeats the business name twice and closes with "Is there anything else I can assist you with today?"

Every word of it is correct. It still reads like a letter from a utility company, and the visitor closes the window before reaching the part that says yes, 8am to 1pm.

Most advice on natural chatbot replies pushes personality: give it a name, add some warmth, sprinkle in a joke. That helps a little. The bigger problem is usually structure, and structure can be fixed without loosening your grip on the facts.

Why accurate replies end up sounding stiff

Stiffness tends to come from a handful of habits, and most of them have nothing to do with vocabulary.

Notice that the fourth one is the only habit that is about tone. The rest are about length and order.

Stiff versus natural, side by side

The quickest way to see the difference is to put real replies next to rewritten ones. These are the kinds of exchanges that come up on a typical service business website.

Visitor asks Stiff reply Natural reply
Do you work Saturdays? Thank you for your enquiry. Our operating hours are Monday to Friday 8am to 6pm and Saturday 8am to 1pm. We are closed on Sundays. Yes, Saturday mornings, 8am to 1pm. Were you hoping to come in this weekend?
How much is a boiler service? We would be delighted to assist. A standard annual boiler service is priced at £95 including VAT, subject to access. A standard service is £95 including VAT. If the boiler is hard to reach it can take longer, so it helps to know roughly where it is.
Can I cancel? Please be advised that cancellations require 24 hours' notice in accordance with our terms and conditions. You can, as long as it is at least 24 hours before the appointment. Want me to pass your cancellation to the team now?

The facts in each pair are identical. The natural version answers first, drops the filler and ends with a question that moves the conversation forward.

Keeping the voice and the facts in separate places

This is the part that stops a friendlier assistant from drifting into invention. Treat tone and facts as two different jobs, stored in two different places.

Facts live in your knowledge base: the pages you crawled, the documents you uploaded, the Q&A pairs you wrote. If a price or policy is not in there, the assistant should say so and take the visitor's details, not improvise something plausible. That rule does not change when you make it sound warmer. The mechanics are covered in more detail in how grounding keeps assistants accurate.

Voice lives in the persona and the welcome message. In SpideyChat that means the assistant's name, a short persona description and the greeting, all set per assistant. A useful persona note is concrete rather than aspirational:

The second version gives the assistant something it can actually follow. It also says nothing about what the prices are, which is exactly right.

One warning. If your website copy is stiff, a grounded assistant will echo some of that stiffness. Rewriting three or four key pages in plain language often does more for the chat than any persona tweak, and your human visitors benefit too.

Follow-up questions that feel like a person listening

A good follow-up question is the single biggest difference between an assistant that sounds like a search box and one that sounds like someone at the front desk. It shows the visitor was understood, and it gathers the detail needed to help properly.

The trick is to ask about the visitor's situation, not about your process.

Visitor: Do you fix leaking taps? Assistant: We do. Is it dripping from the spout, or leaking around the base where it meets the sink? Visitor: Around the base, and there's water in the cupboard. Assistant: That usually means a worn seal or a loose connection underneath. It is worth getting looked at soon so the cupboard does not get damaged. If you leave your postcode and a number, the team can call you back about a visit.

Compare that with "Please provide your name, email address, phone number and postcode so that a member of our team can contact you." Same outcome, very different feeling. One question at a time, each one justified by the previous answer.

A few rules of thumb that hold up well:

  1. Ask no more than one question per reply.
  2. Make the question easy to answer in a few words.
  3. Only ask for contact details once the visitor has a reason to give them.
  4. Stop asking when you already have what the team needs.

For more on the lead capture side of this, see getting emails without killing the conversation.

Where natural goes too far

There is a version of friendly that erodes trust. It is worth knowing the signs so you can pull back.

The honest "I don't have that detail, but I can get someone to answer it" can sound perfectly natural. It is just a short sentence followed by an offer.

What to do this week

Open your conversation history and pick five replies that made you wince. Copy each into a document with the visitor's question above it.

Rewrite each one by hand using the pattern above: answer first, one useful detail, one follow-up question. Then look at what you changed. If you mostly cut preamble, add a line to the persona about leading with the answer. If you mostly softened wording that came from a web page, rewrite that page. If you had to add a fact that was missing, add it to the knowledge base as a Q&A pair.

Run the same five questions again in the demo or your own test window and compare. You are not aiming for a personality. You are aiming for replies a busy person reads in three seconds and trusts.

Frequently asked questions

Can I make my chatbot sound human without it pretending to be a person?
Yes. Plain wording, short answers and relevant follow-up questions feel natural on their own. You can still tell visitors clearly that they are talking to an assistant.
Will a casual tone make the assistant less accurate?
Not if the facts come from your supplied content. Tone lives in the persona and welcome message, while prices, hours and policies stay in the knowledge base.
Should the assistant use emojis and exclamation marks?
Sparingly, and only if your brand already writes that way on the website. Most service businesses sound more trustworthy without them.
How short should chatbot answers be?
Two to four sentences covers most questions. Longer answers suit step-by-step instructions, but even those read better as a short numbered list.

Keep reading

Avoiding the Robotic Feel Without Sacrificing Accuracy · SpideyChat