Cooks and recipes are as important to Artificial Intelligence projects as ingredients and tools
There is no such thing as AI, there are AIs
"Artificial Intelligence" has become one of the main buzzwords for a few months. Every single solution provider or IT service company have been marketing their brand new "AI offering". It can also become quite confusing when one single IT consulting company has several competing internal best AI on the market offerings.
For those who expect a formal course on AI history and evolution over time a lot sites make comprehensive recap.
I will try to give here my 2 cents on all the different types of AI based on project implementation experiences. It can deal with a great variety of tools, of use cases and of course you're invited to challenge and share your own hints.
AI cookbook recipe
The most important thing when talking about AI is in the end not which tool you are choosing (and even less which provider). AI tools are located on different places on the cognition loop : taking raw inputs or refined expertise, producing repetitive outputs or generating knowledge.
Let's consider RPA and chatbots. Cooking the recipe is pretty easy : you can make a wide variety of meals, pretty big volumes, but don't expect to be "Michelin-rated".
Commodity tools like speech or image recognition merely transform raw seeds into usable flour. You would not consider making a mill by yourself and will directly buy the output product. But this is only the starting point to create your cakes, spaghetti (or understand your customer...)
Other tools will require huge investment and refined data as input. Gathering consensus on the appropriate expertise becomes critical to using these tools.
And several non specialized AI tools give you the promise to transform any fresh product into the best dish you'd ever eat... It can be true, but if you go to eat in a 3 stars restaurant, make sure you have enough time and money to afford it...Your company surely have some needs
My chatbot is better than your machine learning
which one of the above examples is "real" AI? This of course is not the good question. For a given use case, there are more or less suitable tools to bring "intelligence" and added value. Let's look at which type of intelligence you are looking for :
- FAQ, service desks : simple chatbots can be of value to bring 24x7 answers to customers and prevent losing time on no-brainer questions
- Repetitive processes (e.g. AP/AR, claims processing) : when talking about a lot of tasks with validation steps with a bit of intelligence to be sure a process is ok, here come RPA (Robotic Process Automation)
- User friendly interactions : When trying to have next gen user interaction, it often comes to voice interactions. I will not challenge here on which use cases this can be relevant or not to me, but will surely do in another post one day. Anyway, building a conversation agent requires a bit more components depending on the level of required interaction. You'll need Speech to Text and probably the opposite (T2S), and a bit of Natural Language Processing
- Recognizing and checking things : This was one of the most complex tools to put in place, deep rooted in the "real" scholar AI. Complex algorithms, a lot of knowledge training, heavy processing power... Thus it comes now as a commodity feature from most cloud vendors. You'll find voice recognition and translation tools, image recognition tools... Of course most of them come via APIs that can be called from anywhere
- Getting more value from customer contacts : The so called "voice of the customer" use cases. In this case, Natural Language Processing and a bit of clustering can be useful
- Fraud detection : For automating rules based on patterns discovery, here come Machine Learning (of course split here are a bit over simplified, as there can be quite some overlaps between machine learning, deep learning, NLP...). Depending on the tool chosen, you will be able to discover the patterns and even directly expose processing based on these results. As an example, Microsoft allows you to create data check APIs in a few clicks
- Pricing calculators, customer service and support : Sometimes you need to apply quite some straight forward rules to a process. in that case, Business Rules Engines can be the real good choice compared to complex custom development. Be careful though on where to use them and how (not talking about licence costs...).
- Contract generation, gathering expertise : When talking about gathering and mimicking real expertise, everybody talks about IBM Watson (coming back to machine learning). Most of the time, this leads to prohibitive costs and in the end inefficiency. Expertise is not unique and sometimes your system will need to deal with contradictions. I've had the opportunity to work with a really interesting french startup in that domain called Khresterion, based on Graph Rule Engine
My Chief Digital Officer wants to "just add an AI block", please help me!
Lets have a look at how to "activate" AI as the next step. Some tools are made to be put aside or on top of what already exists in the IS system. A no brainer chatbot does not need to know anything from your customer, and surely will not be trusted to make complex actions. RPA tools have been designed to be plugged upon existing employees user interfaces.
Most of the time unfortunately, you'll need to have integration with middle and back-office systems. Even the call to external "commodity" cloud systems via APIs require checking legal compliance, GDPR or business strategy. Keep in mind that integrating AI is not (only) a matter of development and IT architecture. As soon as "intelligence" is required, this means :
- Gathering information (either huge amounts of data, or sensitive data). You'll probably need to have data replicates or "big data" systems. A whole data governance and strategy being required in this last case to clearly define boundaries of your data dictionary and data catalogue
- Near real time integration via APIs / services with the company IS systems and third parties. Here again a new set of tooling (API gateways for example) but also of governance in order not to overload mission critical systems
- Not only proposing AI values to end users but also gaining knowledge and expertise from how AI systems are used. A feedback loop is then to be put in place, gathering new data, storing it and crunching it.
OK Google, how to cook "mille-feuille"?
With the same recipe and same ingredients, in the end, what make these two pastries different?
implementing AI solution is of course also a matter of change management and competencies of the individuals on-boarded on the project. It is worth reminding that digital transformation projects, and implementing AI digital capabilities is a kind, is not a matter of IT guys. IT part may be the easiest bit. In my opinion, put 70% of the effort on data, knowledge, business change management... And do the same for constituting your project team...
Keep reading :
Wow! This blog looks exactly like my old one! It's on a completely different topic but it has pretty…
What's up everyone, it's my first pay a quick visit at this web site, and piece of writing is in…