4 Reasons Conversational AI Projects Fail

Post by
Joann Klaas
January 12, 2021
4 Reasons Conversational AI Projects Fail

Not every chatbot and Conversational AI project is a success story. While every company is different and unique in their own way, there are some common pitfalls that should be avoided.

1. Too High Expectations

Our experience has shown that when customers start out with projects, their expectations tend to differ from ours. As a vendor we try to set realistic expectations right away and try to sync with our customers. Going into the project with wrong mindset and expectations will create misunderstandings later on.

One common assumption is that chatbots are easy to make. There is some truth to that. A simple rule-based chatbot that knows answers to some questions, can be made in a matter of hours. However, when developing an NLP backed Conversational AI, what we essentially do is we create virtual a version of a contact center human agent, so in a sense an artificial brain with a problem-solving capability. The training of a real person to become a competent call center agent takes time, the same rules apply to a virtual assistant.

In addition to that, AI has become quite a buzzword over the recent years. Big tech companies are announcing their latest breakthroughs in developing smart AI systems. This might lead persons new to this think that the AI will solve all their problems right away, but as mentioned previously, time and effort needs to be put in in order to get results.

2. Time And Scope Of The Project

Companies and governmental organizations tend to thoroughly plan out their projects, request for proposals (RFPs) and tenders. The more complex and nuanced the initial project, the more time it requires for finalization. As a customer, naturally you want to see results as fast as possible. So companies allocate X amount of time (for example 3 months, which in enterprise world is very short timeframe).  The initial planning takes a good chunk of the project’s total timeframe, which leaves little time for actual development and integration.

We recommend to begin with bare-bones project and initially include features absolutely vital for you. This way you can see if the virtual assistant is right fit for you. As the virtual assistant grows smarter and you gain experience, it is good idea to start implementing new features.

We all want to get our money’s worth. However, when the scope of the project is by no means modest, the allocated budget should be proportionally related to it. Companies tend to originally have 2-3 times smaller budget than the work requested requires. In a case like this, we suggest to go through the initial plan of the project and pick out the feature requests that matter the most. To ease the process, we have developed our own way of finding out where the focus should be directed. We call this method the Topic Rating Matrix (TPM).

TPM enables to screen out commonly occurring problems that could also be solved realistically.

3. Lack Of Communication Between Buyer And Vendor

This might seem quite obvious and overly simplistic view, but compared to traditional software development the Conversational AI is simply by nature more ambiguous. The project is in constant change and the buyer usually does not know what the end result will be or how will they or their customers like it. It is important that the vendor guides the buyer through the process and that the buyer tunes in regularly about their problems. And an always ready support system in terms of knowledge bases, project managers, tech supports and Slack channels is always handy.

To hold virtual assistant related questions in check from your company side, we suggest having one full-time employee as a “Product Owner”. As big companies consist of numerous amount of people and at least as many opinions, it is critical to have one person with the mandate to make decisions. This will make the communication between the vendor and the buyer significantly more efficient.

4. Not Carefully Designed Conversational Flows

It is vital to systematically organize the structure of your virtual assistant from day one. With about 20 topics, it is not an issue, but as the bot grows and the amount of different topics with it, there needs to be a system in place that allows the bot builders and trainers to clearly navigate and modify the actions. If this is not done correctly from the start, the virtual assistant will not perform well later on, which means that extra work has to be done.

The main reason, why virtual assistants are inaccurate in their answers is not because the technology behind it is poor, but because of human error (made by the bot builders, trainers, flow designers). Imagine a scenario, where you have a virtual assistant with 500-1000 topics and the person responsible for the training/building misses one topic and creates a new topic with similar training data. A few individual cases like that will probably not influence the whole system significantly, but with poor organization it is only a matter of time when the AI and the human mind start to make mistakes.