Whether you are using conversational AI for voice or text chat, there are two big challenges that you face in the real world:

  1. The amount of time it takes to build accuracy in intention analysis
  2. Extending the dialog tree to cover all reasonable scenarios

AI practitioners often talk in terms of months, and usually many months, to get to a point where an AI bot can handle 'most' reasonable conversational AI scenarios. The fact that machine learning takes so long normally precludes use of conversational AI for applications other than static servicing applications deployed by large corporate users with deep pockets.

If conversational AI is to deliver on its potential in today's world of short-term projects a different approach is needed:

  1. Limit the scope of what you want to do. Directed questions that steer the conversation are better than open-ended prompts. This is a consistent theme when we at Sytel talk to our partners in the AI community, even with high value AI solutions.
  2. Generate decent baseline data for your intention analysis. If you have recordings of live calls for the business process and can use ASR to transcribe, this gives you something that you can use to train Naïve Bayes filtering to give a set of keywords that will be better at getting to acceptable accuracy quickly than a 'made up' baseline and lots of machine learning.

And there is more... Make sure that break out to a human agent is seamless, quick and provides the agent with all the context of the conversation. This requires your vendor to provide technology and tools that you won't find in the average cloud ACD offering.

To make the customer experience of transfer from AI bot to human agents seamless and happy one, agents must be ready and waiting to deal with breakout from the AI bot. SLA time for this is actually just a few seconds, not 2 minutes! And factor in that, in the real world, you can't keep a pool of agents idle 'just in case'.

What the ACD must know is

  • how far all of the bot sessions have progressed through the dialog
  • the likelihood of dropout over the next seconds
  • how many agents to hold back from regular queue activity, so that they can be there to deal with projected drop out from the AI bot.

In an earlier blog, we suggested that this circumstance would apply only to outbound bot activity. But it applies to inbound as well. For many of the prospects and customers being moved to live agents have probably had their patience well-tested by the bot and will be in no mood to stay long in an inbound queue.

The solution lies in the same kind of AI technology Sytel has developed for optimising predictive dialing behaviour with live agents.