Current technologies: an overview
Definitions vary, and rather than getting caught up in semantics, I use the common descriptions to reflect the broad areas.
Artificial Intelligence (AI)
AI is a term to reflect a series of instructions of how to go about something and the end goal, without specifying what to do in every circumstance. The instructions or algorithms are applied to a specific data set to achieve an outcome. A parallel from childhood education could be considered the difference between responding through ‘understanding’ of a subject area (AI) and learning an answer by rote (rules based responses).
In the context of sales bots, AI represents a means of responding within a conversational message flow (or dialogue) that is either an addition to or an alternative for search and filter: intended to give a more focused or complete response to a question or requirement.
Some purists consider certain bots are not AI, but remember AI is merely the name for a machine taking a series of learnt facts or rules and applying these to a situation to define an outcome. What matters is how suitable the AI engine is for the purpose to which it is to be put. There is no better or worse AI engine, but there are more or less suitable ones for a given purpose.
Natural Language Processing (NLP)
NLP is a mechanism of taking the actual words used by a person and ascribing a defined semantic meaning or intent to the words, which may or may not be a literal interpretation. The first key benefit of NLP is that its use should feel natural – use your own words to describe what you want. The second key benefit is that it can be used in many situations – in fact pretty much any form of verbal media – but most typically in written messaging media such as instant messaging, live chat or email.
NLP is best used for handling a narrow range of requests even though you can ask in many ways. For SMEs this is a really key point – you need to be doing the same thing many times for NLP to really work for you. The alternative: prompt based approaches are undoubtedly more accurate and might be worth considering.
Conversational Commerce (CC)
NLP brings us nicely to CC. The big hype about CC is because consumers are moving. Less time is now spent in apps and more time (now the majority) in messaging platforms. Instead of doing something in an app, or broadcasting using a medium such as Twitter, people are spending much more time in dialogue person to person or in a group. It is therefore important to those that want to get access to some of that time and attention.
Conversational commerce is the ability to reach and sell to a customer whilst in a conversational medium such as messenger or chat.
However, there are some qualifiers as to who should be interested. First up – are you selling direct to consumers or are you selling to businesses. Second – is it really practical to market and sell your product without reference to some form of catalogue or menu?
CC is often confused with mobile. Mobile merely reflects the device that your content is being accessed from. Granted the phone is the ‘first screen’ for personal use for most. But that’s just the point. For personal use. Before you get into mobile, understand the nature of your addressable market, and if it's B2B when is your prime selling time? A user at work in an office for example is better reached through a desktop or laptop. What other applications are they using (including browser accessed, cloud based services)?
Processing approach: Probabilistic, decision trees and method rules.
Let's move on to the approach of finding an answer (or even deciphering the question).
Unless you have very simple needs, be cautious of simple question/answer technologies, which are little more than FAQ responders. Some systems combine several question:response pairs but process them independently. Whilst apparently more sophisticated, these selectors tend to rigidly funnel an answer, ignoring ambiguity, even if they allow individual responses to be changed. Combining multiple question:response pairs before answering means you can deal with more complex queries and provide advice. This is the equivalent of a salesperson gathering all of the customer requirements before proceeding to the next stage of a sale.
Whether the system deals with one or several inputs, they are then processed to find a response.
Probabilistic approaches look to match patterns – identify the input pattern and provide the specified output. They work well when there are many hundreds or thousands of repetitions of very similar transactions. Fed with a history of incidents and either the associated ‘correct responses’ or method to create one, this approach provide the most likely answer. Self driving cars use this method to predict what a human driver would be most likely to do in a certain situation.
If there are insufficient repetitions to be viable, or there is zero tolerance to errors you may be better served looking elsewhere. A decision tree could be a good option to pursue. However, you are getting closer to providing a specific response for each specific question. The training could be laborious, but the outcome will be highly predictable – great for regulated industries.
The last main alternative is a happy medium: the method rules based system. These are just what they say: teach the method rules and let it handle the enquiry in the most optimised manner drawing upon the history of similar interactions.
Technologies can be delivered as part of a web page, as an unconnected window on a page, or as connected window or overlay. Building into a page means changing each page that accommodates the bot. Using unconnected bots restricts ‘show me’ functions (which is also a weakness with most chat applications). Connected but separate deployment can give richer functionality including concierge style experience.