Recently, Converse.AI participated in Bot World, a dedicated U.K. event that brought together brands, developers, and pioneers of the hottest tech trend.

We hosted and moderated a panel session featuring Facebook, Sage, and Twilio on “Top considerations when building a relevant bot.”

The conversation was engaging and thoughtful, and I wanted to share the most important highlights for anyone thinking of building a bot.

What should I avoid doing?

The answer to this one was clear: Don’t spend 12 months building something that you think is perfect — it won’t be. The nature of a conversational interface is such that you will be constantly surprised how humans interact with your bot. No matter how long you spend building and training your bot, users will always react differently than you expect, and someone will always be able to trip your bot up. And that’s OK.

The beauty of using a bot is that even if the bot doesn’t work properly, you can still get clear feedback on how it’s gone wrong, and bots can be easily changed in minutes in response to feedback.

“The agility cycle is really fast: much faster than apps,” says Jon Heaton, a Partner Engineer at Facebook.

What should I make sure I do?

While you are iterating and tweaking your bot, it’s still really important to have a human agent available when necessary. Bots are great at assisting with communication for businesses but they should still be able to integrate a human in the loop to make sure the user experience is as smooth as possible.

“You and your employees are the only ones that know your customer journey completely, so use that knowledge and experience,” says Chris Book, a marketing manager at Twilio.

Whether users reach a manual call center or are messaging an automated bot, both bots and humans will have certain processes to go through or hoops to jump through. While the phrase “Computer says no” became a comedy catchphrase, there are limits to what a bot is able to do on its own, so make sure that the user’s expectations are managed. Only humans will be able to apply common sense, though it can be taught to the bot when you are training it.

“You have to set expectations on its capabilities and limitations the very first time someone uses the bot,” says Kriti Sharma, a mobile product manager at Sage.

So the overall advice here is to start small, iterate, and grow. Consider your users’ expectations at every stage and make sure you are taking them into account with every use of the bot.

NLP or buttons — which is better?

When Facebook launched its bot manifesto at F8 back in April, it became clear very quickly that people were not grasping the full concept of a conversational interface.

There are now over 30,000 bots on Facebook Messenger, making it by far the most popular platform for building bots. Facebook has rolled out many features that enable people to use a combination of buttons and text-based features, as well as natural language processing (NLP). The advice given by the panel was to use a combination of these to make sure the users familiarize themselves with the bot interface as quickly as possible.

As with every step of building your bot, the key is building with your users in mind, while keeping the overall desired outcome of the bot front and center.

NLP is useful, buttons are useful, but certain scenarios will suit different approaches better. Long-form text bots will require NLP, whereas simple question bots, like those for ordering a pizza, may benefit from more buttons. Of course, a combination of these is most likely, and you will almost always need to handle some language input, such as addresses for pizza deliveries or notes to the chef.

Integrations with third parties can also help with adoption. For example, an integration with PayPal or Stripe to take payments can help users feel confident and secure when sending card payments through a bot. It’s not just about the features that you include; functionality and usability are really important, as is user confidence.

What does success look like?

All of the considerations that were discussed last week focused on the customer experience and adoption curve. The message was that the most important part of your business is your customers and their experience interacting with you. Your top considerations when building a bot shouldn’t be the technological challenges. These have almost all been solved for you. Choose a bot-building platform that makes it possible for you to build something your customers will love and that will also take care of the innovation and enhancements for you.

The main focus for you as a bot builder should be making sure the personality and overall usability of your bot are right. To do this, you need to consider how you are going to add context and personality that your end user will respond well to. All of this ensures you are creating a friendly conversation for the user, who will then most likely return to use your bot in the future.

“We all have a responsibility to rid the world of crappy bots!” says Sharma.

Outbound interaction with customers at specific points of their engagement cycle is not new. It’s not just responding to inbound messages but also about being able to push messages to customers at the right time, with the right message.

Bots can be used to initiate conversations based on previously collected data. This could mean offering a taxi when you know someone has landed after a flight they booked through your bot, or asking about upgrades or extended stays when they have booked a hotel. The key is to make sure the message is relevant and welcomed by the end user. And don’t forget, you may well still need human escalation for those outbound messages.

Getting the balance right when adding value through messaging and avoiding “spamapocalypse” is the true test of a successful bot.