‍

Can you come up with a chatbot service you actually love? Hold your thought: you can’t answer with a multi-billion dollar funded service backed up by an army of humans (meaning Google Assistant, Siri or Alexa are ruled out). Yes, it’s hard to find live chatbot services that are enjoyable to interact with. After a decade of inflated expectations, high-quality and accurate chatbots are still a rare thing. No wonder why Gartner places chatbots, jointly with NLP, in the trough of disillusionment since 2021. It’s a pity to acknowledge chatbots are synonyms of crappy service in most people’s minds. The technology simply couldn’t live up to end-users expectations.

‍

‍

Where did it go wrong?

Technology is not the only culprit. The market is flooded with low-entry chatbot solutions that make it very easy for anyone to launch a bot in a couple of clicks. These plug-and-play black-box solutions hide the crux of chatbot complexity and deceptively make developers feel happy for releasing short-sighted and non-production-ready services that would better qualify as betas or demos.

At WITH, we learned the hard way that operating chatbot services in live environments while maintaining accuracy is the real challenge. We launched a number of customer service chatbots with our own technology. Some are still alive and do qualify as rare things. We discovered that once your bot deals with real end-users, new difficulties unfold:

• End-users don’t make short statements. They enjoy telling their life, with complex intentions full of insignificant details. All of a sudden your chatbot training dataset appears terribly weak and irrelevant.
• People don’t write clean sentences. They make a bunch of grammatical or orthographic mistakes mixed with abbreviations. Now your NLP engine starts to spin in the void.
• They often ask unresponded questions, simply because there are no actual responses known to the bot. The lack of ability to adapt your knowledge database to users' needs ruins your service.
• Users heavily use a vocabulary specific to your business domain. It’s not about entity extraction but business context awareness. The biggest language training datasets you can find won’t understand what your business is about.

This is the point when the service rapidly degrades. You can’t seem to pass the initial training bootstrap chasm, you have no visibility over your AI blackbox quality and you can’t measure the impact of playing with all your AI knobs. Eventually your accuracy quickly becomes asymptotic with a ceiling you can’t break through (our best implementation reaches 60% over 10k real questions). It’s time to acknowledge you are desperately missing the right tools to operate the beast. Two outcomes are available to you: you give up or you rethink how to address correctly the problem you are facing.

Fixing the mess with 1.2M€ open-source software

WITH bet on the latter because there is hope: after the trough of disillusionment comes the slope of enlightenment (according to Gartner). This is why we are investing more than $1.2M to build a new generation of FAQ chatbots that comes with “operational excellence by design”.

Tooling and operational capabilities will be at the core of the product to give control over the maintainability and adaptability under live conditions. Because too many “magical solutions” are hidden behind rigid black boxes, we are doing this with an open-source model. Our aim is to bring to the developer's community a new powerful tool to better serve their projects when it comes to conversational experiences and NLP engines against FAQ knowledge databases.

We will deliver this new exciting product within an 18 month project, with a team of 10 outstanding persons eager to bring innovation to the space. We will communicate our achievements regularly, push our releases on a git repo, and give visibility over the upcoming roadmap. Hopefully our efforts will grow a new chatbot community willing to fulfill the desire of keeping their service live rather than dying in the attempt.

Stay tuned

‍