Cracked Conversations: What to Do When Chatbots Aren’t Enough

Conversational AI, or chatbot, vendors, are everywhere, deafening customers with the promise of AI-Powered solutions for their customer service needs. According to Capterra, 158 companies currently offer chatbot software. In Forrester’s evaluation of the emerging market for conversational AI for customer service for Q2 2019, the analyst firm identified the 14 most significant providers in the category – [24]7.ai, Avaamo, Cognigy, eGain, Indenta Technologies, Interactions, IPsoft, Kore.ai, LogMeIn, Nuance Communications, Omilia, Saleforce and Verint.

This makes understanding what works best to improve customer experience hard.

Chatbots work best guiding users along straightforward, well-defined conversational paths. If a customer asks new, unpredicted questions the typical chatbot gets confused. More complex questions require complementary solutions.

Sinequa offers one such complementary solution – Enterprise Search that can work with chatbots to help customers and employees find what they need.

We have spoken with a number of companies ranging from those considering the technology, to building prototypes, to deploying chatbots in customer-facing applications.

Several of the concerns about the value produced by chatbot deployments

• Slow Conversation speeds
• Conversation path-sets grow larger and longer
• Low accuracy because the chatbot was unable to answer and was unable to maintain the chat
• High development effort with too many expert hours spent conceiving, designing, deploying, and maintaining those conversational paths.

Some Reasons Why?

Chatbots work best when guiding a well-defined type of user through a set of preconceived conversational paths.

The typical chatbot’s tooling provides a graphical interface, and some testing capabilities; conceiving, designing, deploying, and maintaining those conversational paths will be up to you.

• When you consider how many paths a user might take, multiplied by the number of user types, it can grow to an astonishing amount of work.
• When chatbots have a lot of this work to do, they tend to slow down compromising, the chat experience.
• Most requests for information are ‘ad-hoc’ and therefore not well-suited for a pre-planned and pre-built conversation flow.

When Do Chatbots Make Sense?

• An example is a chatbot at your local bank. E.g. checking, savings, mortgages, lines of credit.
• Those offerings have a limited number of actions. Checking deposit, transfer, bill pay, balance inquiry.
• The site is often for reference, not as much for execution. To actually open up an account type, you typically have to apply in-person.

If you can’t narrow the scope to specific user-types and paths like these, then the outcome of multi-step “chats” is by definition, less predictable, leading to a higher failure rate.

This also makes it difficult for some chatbots to get a PTO (Permit to Operate), because companies have not let applications go into production that couldn’t guarantee outcomes. This is to avoid “Rogue AI” situations, among other things.

Addressing the Challenge

Enterprise Search, like Sinequa’s, leverages natural language processing (NLP) to get users the most relevant content, without the chatbot’s requirement that the conversational path be designed, built and maintained.

Where chatbot interactions are sometimes helpful, that chatbot can connect to enterprise search; when the chatbot gets a user’s request for information, the chatbot can refine and forward the request to the underlying Sinequa search, then channel the results back to the user’s conversation.