Business Readiness for Conversational AI
What is Business Readiness?
Any conversational AI system is driven by two core components, the business or functional requirement and the IT configuration or build.
Since the AI system in this case is conversational, (i.e. it replicates the human the human conversation), the understanding of business must be built in a way, that they could relate to it in human terms. E.g. — Biasing is a common problem in classification AI systems. Biasing happens when data instances for one label are significantly higher than another label which results in AI picking up label with high data records count more often than the other label(s).
This can be explained to business as follows:
As a human we’re geared more towards a specific skill which we have picked up over a period of time. A person who is educated in mathematics is more likely to solve complex mathematics problems and a historian can shed more light on the historical aspect of an artifact. A mathematician would be relatively less accurate while dealing with history than a historian. This is a very common example of biasing, where we are geared towards one area(label) and are less accurate towards others. We, humans, also have, to a certain extent, a level of bias in us.
Before we start a journey on AI, we must educate and get the business ready for the challenges and requirements that will come in the path.
Business readiness can be broadly classified into following categories:
- Understanding the AI
- Understanding the data
- AI principles training
- High-level requirement finalization
- Data gathering and preparation
How to get business ready?
Business readiness should be done in three steps:
- Training the business on AI and principles
- Understanding the data and requirements together with business
- Data preparation
Training the business on AI and principles: — Business should be given sessions on AI and how an AI system work based on data. This session needs to be strictly high-level, touching only upon the principles and should not be taken to technical details. A good example for something that can be explained to business in this session is “what is an intent”. A one-liner explanation for this is, anything that users wants to perform or get information for is an intent.
Get Shashikant Jha’s stories in your inbox
Join Medium for free to get updates from this writer.
Understanding the data and requirements together with business: — Once the business is comfortable with the AI system and its inculcation in existing process, the requirements should be understood. The requirements discussed at this stage should be a high-level overview and expectations from the AI system. They shouldn’t be broken down to sub-parts.
E.g. — The high-level requirement for putting a conversational AI system as a VoiceBot(IVR + Chatbot)would be CONTAINMENT and 24x7 availability. We shouldn’t drill down the requirements at this stage. Instead we should focus on historical data to understand more on how these goals could be achieved.
While we’re in this stage, the reporting KPIs should be defined. We must measure success of the entire system as well as individual components.
To elaborate, if a system has more than one AI component joined in series, the failures from one system would propagate to another system. This must be accounted for while defining the success of an entire system. The success KPI defined must align with the end-goal of the system.
Data preparation: — Once the requirements and data are understood, a master sheet should be created for data. The master sheet must contain the following details and should be a living document rather than one-time effort.
- Intents
- Purpose of intents
- Trigger condition of intents
- Data (Training phrases) for intent. Since this requires some analysis and principles to be followed as to how many training phrases should be there per intent and in all, this must be done with IT in a working session.
There should be enough data after this stage for training and validating the AI system. The validation set created must be different from the data being used to train the AI system and the content of data should be increased if needed.
The test set should never reach developers, as the visibility of test data could lead developers to do a skewed training of system and could effectively render the whole process useless. The test set must always reside with business, and if there are failures while running the test set against the AI system, only the failures (preferably patterns) should be communicated to developers.
What to expect from business readiness?
Once the business is ready to embark on the journey, we will have created the below listed documents for future reference:
- Master sheet of intents defining the purpose, trigger condition and training phrases.
- Clear understanding of end-goals and agreement on how to achieve it.
- KPIs defined and agreed upon for reporting and measuring success.
If you’re interested in Conversational AI, you can reach me on LinkedIn or mail me at jhashashi669@gmail.com
