How to ensure your AI project goes from POC to Production?

As the year turns to 2020 and we move forward, so does the advancement in AI and Machine learning algorithms, leading to even bolder and bigger prediction for 2020. In its simplest terms AI or Artificial Intelligence is the intelligence of a machine that can sense surrounding environments and make decisions to adjust accordingly without (or minimal) human intervention.

Statistically 47% industries or businesses have implemented AI in at least one of their operational functions while another 30% are piloting AI. Business processes across industries are being transformed using AI and machine learning, and the adoption is going to be accelerated only if the investment by big tech giants in the field is any indication. However, as Gartner report suggests the biggest challenge in AI implementation is not data availability or quality. The biggest challenge is that 80% of project never sees the light of production. It can be tied to many things but having an unreasonable expectation without understanding the complete foundational lifecycle of an AI project is the prime reason. Taking AI from POC(proof of concept) to production is what differentiates men/women from the boys/girls. How an AI project is planned from its inception can tell you the future course it is going to take. So here are few critical aspects you want to ensure that your AI team has covered and understood well

• Which part of use case case or solution is “actually” AI? To take a leaf from Batman’s book and customize it a bit for AI - “With great technological potential comes greater hype and greater need to balance”. AI has piqued so much interest in recent times that there are many impersonator solutions which simply claim itself as AI but has no element of AI in it, at max it is a rudimentary automation solution. A true AI solution has two important elements - How does the system try to mimic human brain and, more importantly, how does it learn over time. If you do not have sufficient and good quality data to build a machine learning model, there is a good chance you will start with a heuristic model. Data collected through this heuristic model will be the bed-rock of future AI model. This should be understood clearly and communicated precisely. Ask your team or the service partner which specific part of the solution has AI and how?
• Do you have data? If yes, where? If no, how will you get it? Plan for data - AI solution feeds on data, no learning will happen without a data. This is why it is not surprising to see ~60% time and effort of AI project going into data engineering and preparation. A knowledgeable team will be able to think about how the data will be acquired, what is the quality of existing data and more importantly how future data acquisition will mature the model. Wrong data steps might result in biased model, inaccurate model or at worse completely negate your hypothesis you started with. Questions such as - What kind of data do I need for the solution to function well? How much data do I need? Where do I need to source the data from? Does it have to come from internal systems or can we source it from third parties? - need to be explored in detail. An experienced AI team will tell you if you don’t have enough good data to help an AI solution make strong predictions, and if a workaround (like using public data) is available. 
• Where is the inaccuracy management plan? - As we see implementing AI projects across industries and talking to executives planning AI pilot projects, major focus, rightly so, is given on building and improving the accuracy of the machine learning model. However, an equal (if not more) important aspect is to have a plan for managing inaccuracy of the model. No matter what you do, machine learning model will never be 100% accurate, period. Even the best data and model will have some level of inaccuracy, that is what the fundamental theory of any probabilistic model dictates. There is a very good chance that your AI model will start with a lower level of accuracy than anticipated. Only way to have a 99% accurate model in production is to have 70% accurate model in production. You have to push lower accurate model in production so that real-time data can mature model to the desired level. Of course, you have to do this in a very controlled environment without risking the whole project. A robust inaccuracy management plan can fast track your production roll-outs which helps businesses realize its goals faster and additionally provides more data which in turn enables accuracy improvements.
• Human and ethical side of AI - Ethics and human aspect of AI can not be corrected in hindsight, it has to be the foundational block of AI implementation plan. Let’s talk about biases first - machine learns biases from its learning dataset. There are techniques available to eliminate/minimize biases and those techniques should be incorporated in the very foundation before models get to feed on data. Secondly - keeping humans in loop and transforming workflow accordingly should be done in non-frictional way. For example - while planning for building a chatbot for customer support, at what point you transfer the control from bot to human and vice versa? Whether you should do this bot-to-human handshake in completely transparent way and let your user know that agent switch has happened in background or will that create friction? A capable implementation team will consider human and ethical aspects of AI at every step of project from data collection to data preparation to model building to operationalize.

Thanks for reading. I welcome your thoughts, feedback, agreement and off course disagreement.