Writing for Chatbots


When I wrote an article a few weeks ago about conversational web journeys being the future, I had no idea I’d be pulled into a chatbot project just a few days later!  It was the ideal opportunity to experiment with writing for a completely different medium.

This was an A/B test where we had to keep the same question set as in the existing quote journey – but adapt it considerably for a chat version.  But there was a lot to consider. Here are some of the key things I learnt along the way.

Tone of voice

Firstly we had to think about how this brand would speak to a customer.  Certain colloquial phrases were out of the question as this brand had a more formal style.  For example, we’d need to use ‘Thank you’ rather than ‘Cheers’ and ‘Good afternoon’ instead of ‘Hi.’  

It’s always important to stay true and relevant to the brand’s target audience and not just assume a new persona because you’re in a chatbot.  I found it helpful to think of a person – if your brand was a person, who would they be, and how do they speak?

If I was doing a similar project again I’d map out all the standard phrases for this ‘person’ as a kind of conversational style guide, before starting any content creation.  

Lack of help

In our standard quote journey we have ‘help’ copy for questions that customers frequently get stuck on.  In a chat bot you have to approach help slightly differently.  The purpose of a chatbot is to feel natural and simple, so each bit of dialogue shouldn’t really be more than a couple of sentences.  We had to make the questions as easy to understand as possible without the need for help!


In chat, it wouldn’t feel natural to just fire questions one after the other without acknowledging each answer provided by the user.  You can use this as an opportunity to add in a human element, like a compliment such as ‘that’s a nice name’ or you can just use a simple, ‘Thank you, that’s great.’  We used a combination, also even playing back the answer in some instances, to reassure the customer we’d heard them!

First name terms

It was important to capture the customer’s name quickly, so that we could reference it throughout the conversation.  This is different to a standard web journey when customers tend to not want to give out personal information too early on.  Our research has shown customers can feel anxious about how it could be used when they’re not yet committed to buy.  We took it a step further by also asking the customer how they’d like to be addressed through the journey.

Limit responses

With a chatbot you need to make sure the conversation is kept on track.  As soon as a user starts to deviate you’re at risk of not having an appropriate response.  For some questions we made the answer multiple choice rather than free text entry to limit that scenario.  We also prepared some standard responses to the free text answers to ensure if the user didn’t give an appropriate response we could steer the conversation back on track.  This bot wouldn’t be intelligent enough to start learning and responding so all dialogue from the bot had to be pre-prepared.


We started out by using Post-Its to produce the conversation, and identified the ‘happy path’ (the version where we get exactly the responses we want), and then looked at the deviations.  We then mapped this into a spreadsheet so that the ‘rules’ could be documented beside each text entry.

Once the bot prototype was built we could then test and iterate, as well as identify any scenarios we’d missed previously in our mapping.  Testing also helped to try different responses and check that the conversation still hung together and felt natural.

The bot’s not live yet, but as soon as it is I’ll update this link so you can have a go yourself!