The AI Bible
The AI Bible includes all content needed in order to successfully complete the ML Essentials Assessment and become an industry acknowledged certified AI business implementation analyst. We provide all content free of charge, so everyone (certified or not) can benefit and has access to knowledge that helps accelerating AI innovation.
The Bible covers the following topics:
History and Definition
Machine Learning Definitions
Eight different Machine Learning methods
Deep Learning / Artificial Neural Networks
Data Security & Ethics
AIcompany Talent Framework (ATF)
AIcompany Pilot Management Framework (APMF)
AIcompany Pilot Management Framework (APMF)
In previous chapters we have talked about all major knowledge areas of machine learning, and now we will be talking about creating real business value with the help of the 6 steps that help you start machine learning pilots.
AIcompany designed a framework to help managers get a quick-scan selection of all possible processes where Artificial Intelligence would have a great impact. Ideally, we look for projects and models that would be able to make a difference and deliver business value within 6 months or less (in production).
Confirm the highest priority opportunities for AI within the organisation (step 1,2)
Identify the bottlenecks (step 3,4)
Resolve the conditions and pilot-set up (step 5,6)
Step one: Client First?
We see organisations starting machine learning initiative because they see an opportunity to generate more revenue (Return on investment driven). The APMF requires you to start identifying from a customer satisfaction point of view. Not only will it strengthen the business case but will also help you find support within the organisation. Therefore, write down three (maximum) or less Client First themes where an increase of customer satisfaction (i.e. NPS scores) will have great impact (Core Business). These themes should be described however your customers /clients perceive the full service / product. For example, if you sell mortgages, please remember that customers are not looking for mortgages, most likely, customers would much rather get a house without the burden of a mortgage. A question that might help you to come u with customer satisfaction themes; ‘Where would our customer really appreciate a more proactive approach?’
Please list the following answers for each Client First Theme:
- Activities that consume significant resources
- What are currently painful points this Client First Theme. Please plot all of these points on a four matrix: (Value) High-Low and (Time) Short term – Long term.
- Would this change without this project taking place? (Especially the High & Long Term pain points)
- Are best practice examples (per pain point) available?
Step two: What must be done manually and repeats itself often?
Answering this question would in most organisations result in a long list of activities and processes. Please try to group and label them, keep it simple (measured in time, hours for example, %).
As mentioned, search for procedures and decisions that are made frequently and consistently. We recommend businesses to start looking for suitable core processes. Try to collect as much data as possible about how the decision was made. In the example of the loan approval, what customer data was important for approving the loan? (Who made it? At what time of day? How confident did they feel in the decision?).
Step three: What do we actually know?
Very important, identify what data is available and is currently stored. Can data be used to solve it? What are the key sources of data within your organisation? Similar data sets might have been stored in different formatted labels or back office systems. Please identify;
Are processes or activities labeled?
How much data is stored?
What Meta data is available (data about the data)
Is the data structured or unstructured?
Is data linked to results?
Is available data accurate and complete?
Are there any data custodians? Are they really responsible for their data?
All of these questions will help you determine what kind of model to use and how to make it more accurate. Need inspiration for useful data sources, please check;
Product listings: category, SKU (stock keeping unit), prices, stocks, margins, sales history
Customer listings: purchase history, LTV (loan to value), loyalty card information, email, physical addresses, online profiles.
Suppliers: name, location, products, commercials, distributors, restrictions.
Distribution: warehouse operating information, delivery.
Sales channels: store location, competitors, sales, staff, desktop/mobile
Paid marketing: ad words, copy, bid prices, clicks, display schedule, landing page.
Alternative data sources: social media accounts, weather forecasts, consumer surveys (kantar), Trustpilot, Experian, news sites.
Finally, we recommend you to align all answers and insights of 1,2, and 3. Preferable plot all of these actionable outcomes into a feasibility matrix / model (suitable for your specific organisation). Aligning all of these insights should help you tackle the right problem!
Step four: Design your model
Step four is considered the most time consuming step. Here it comes down to the creativity and skill set of the AI specialist / Data Scientist you are working with in order to deliver value. What could an entire dedicated team of the best employees do for this one special client? What data (selection) would be required to design an appropriate learning model that will have a high accuracy performance? Design a model that is able to predict, handle, plan, advise and support this delivery. Remember, finding the perfect fit requires time and many Trial-and-error’s. The pitfall we see often encounter in real world organisations is a project team that started out with a highly complex model to solve a problem that can be easily simplified and solved by a relative simple model.
Unsupervised machine learning might be the end goal (agents), but try to combine methods. For example, while doing market research you might want to segment consumer groups to target specific website behaviours, a clustering algorithm will be sufficient. Apply clustering techniques to derive smaller number of features and use those features as inputs for training a (more simple) classifier model. Image 4 shows a machine learning workflow we often see in real world organisations.
Step Five: Select your Dreamteam
(See AIcompany Talent Framework for details, previous chapter) Please do not hesitate or be shy in creating the pilot Dreamteam. In larger organisations we noticed that even a one-hour commitment from the right domain specialist could make a great difference. Selecting your team is also a great way to create ambassadors across several departments and divisions.
Step Six: Pilot Time
The pilot should demonstrate what a successful implementation of a machine learning model would look like. Deployment in a production system is recommended. How accurate is the machine learning model performing to new real world data? During the pilot phase there should be more than enough data and feedback from end-users available that allows you to fine-tune the business case, and rethink return on investments for further scaling.
Good luck! Please do not hesitate to contact us if you would like to have more information or want to see practical cases (in your industry) where APMF designs were implemented.
AIcompany Talent Framework (ATF)
The team should consist of 4 up to 6 members. Our guideline; “A team should not have more members than you are able to share a pizza with”. Having shared quite a number of pizzas in respectively effective teams, we kept this guideline. A common mistake is assuming that a highly technical stack is needed. We broke the requirements back to 4 capabilities.
AI Business Analyst
The AI business analyst has enough domain and business expertise in order to understand business behaviours and industry specific knowledge (Business Intelligence). The AI business analyst usually has frequent contact with customers and but is also curious about data. He or she does not necessarily cover technical domains but has to be able to use data to open up new product, business, and revenue opportunities.
Data Scientist/AI Specialist
The Data Scientist builds the model and algorithms. To specify, there is a difference between AI specialists and Data Scientists in terms of research capabilities. Across industries the job description ‘Data Scientist’ is lately used for all kinds of tasks related to data.
We believe this makes it harder for employers to distinguish these specialist capabilities. We see AI specialists (e.g. Deep Learning Scientists) that are able to understand and adjust algorithms and taking the smallest details into account, whereas general Data Scientists are in some cases limited to applying proven methods.
The Programmer has all of the computer science fundamentals and is familiar with solution design. Relational algebra is important to create effective solutions and interactions with existing IT landscapes and architecture (i.e. prepare for production). Another perk would be coding capabilities such as SQL and SAS in order to understand databases and infrastructure. Integration examples; including third party’s software packages (API’s) and creating and setting up features in a mobile app.
The Chief is responsible for creating a hospitable environment, knows the organisational culture, different departments, policies, guidelines, back-office systems, and funding streams. In simple terms, ensures that all stakeholders are up and running (in an Agile working environment this includes Scaled Agile Alignments, Product Owner roles) and is able to estimate the time and resources that will be required to get it done. The chief is not satisfied with simply solving the problem, but makes sure that the solution is accepted, adopted, and applied by the right people so that efforts will have a great impact on the organisation. This chief might not be an expert on tools that are used by the rest of the team but will have a good (helicopter view) understanding of all the activities. The Chief is preferably someone with a (relative) higher position within the organisation.
Develop your own talent:
Is your organisation lacking good Data Scientists and are you looking for ways to develop good data science employees? AIcompany noticed that the most frequently entry routes to data science are:
Science, especially physics and math