“What is the business problem that we are trying to solve here?”

The most important phase in any software project is to understand the business problem and create requirements. ML-based software is no different here. The initial step includes a thorough study of business problems and requirements. These requirements are translated into the model objectives and the model outputs. Possible errors and minimum success for launching need to be specified. The most useful question to continue working on the AI/ML solution is “how costly are wrong predictions?” Answering that question will define the feasibility of the ML project.

Work Flow Decomposition

Each task of the entire business process needs to be decomposed into its constituent elements in order to see where prediction (ML model) can be introduced.

The Workflow Decomposition

To anwer the question “how to implement AI/ML”, we follow the next steps:

  1. Identify the concrete process that might be powered by AI/ML (see the Figure above).
  2. Decompose that process into a directed graph of tasks.
  3. Identify where humans can be removed from the task, meaning, what task can be replaced by a prediction element such as ML model?
  4. Estime the ROI for implementing an AI/ML tool to perform each task.
  5. Rank-order the AI/ML implementation for each task in terms of ROI.
  6. Start from the top of the list and structure the AI/ML implementation by completing either the AI Canvas or the Machine Learning Canvas.

The AI Canvas or its alternative, the Machine Learning Canvas, assist and help to structure the breakdown process. They also help to articulate exactly what is needed to predict and how we react on errors made by the prediction algorithm.

AI Canvas

The AI Canvas was proposed by A. Agrawal et. al in their book “Prediction Machines. The Simple Economics of Artificial Intelligence.” 2018, and “is an aid for contemplating, building, and assessing AI tools”. The example of such canvas and the description of each component is provided in the Figure below:

AI Canvas by Agrawal et.al

Figure Source

Machine Learning Canvas

While the above AI canvas represents a high-level structure of the ML/AI implementation, at some point we would like to specify both the vision for the ML system and the specifics of the system. To achieve those goals there is another tool, the Machine Learning Canvas, as suggested by Louis Dorard. This canvas structures the ML project and helps to specify the core requirements to realise the project. Initially, we identify the objective by answering a question what do we want to achieve for the end-users of the predictive system? Next, we connect the business goal to the ML task.

The central part of the canvas is the Value Proposition building block, which describes products or services that create some value for customers. Typically, we answer the following questions: What problems are we trying to solve? Why is it important? Who is the end-user of our system? What value does the ML project deliver to the end-user? How will they use your outputs/predictions?

The remaining canvas is divided into three broad categories: Learning, Prediction, and Evaluation. The Learning category is responsible to specify how the ML model will be learned. The Prediction part describes how the prediction is performed. Finally, the Evaluation category contains methods and metrics for the ML model and the system evaluation. The following machine learning canvas is an example provided by Louis Dorard:

Machine Learning Canvas

In total, the Machine Learning Canvas is structured as ten compound blocks, such as Value Proposition, Data Sources, ML Task, Features (Engineering), Offline Evaluation, Decisions, Making Predictions, Collecting Data, Building Models, and Live Evaluation and Monitoring. Each of those blocks is focused on one aspect of the future ML application:

Value Proposition

This is the crucial blocks in the whole canvas. Here we should answer three important questions:

  1. What is the problem? What objective are we serving? What are we trying to do for the end-user?
  2. Why is it important?
  3. Who is the end-user? Can we specify the persona?

Narrowing the domain of the problem could be useful for the next question regarding the required data. For example, instead of creating a universal chat-bot, build a bot that helps with scheduling conference-calls.

Data Sources

Data is essential for training ML models. In this block, we clarify all available and possible data sources to be used for the ML task. As an example, we might consider using:

Furthermore, we should clarify the hidden costs of a machine learning application.

ML Task

After clarifying what data is available, we brainstorm what type of ML should be used. Here are some examples of questions that might clarify the ML Task:

Features (Engineering)

As every ML algorithm requires input data in the form of features, we should clarify how should the input data be represented.

Offline Evaluation

Before any implementation of the ML model training, we would need to specify and set up the methods and metrics to evaluate the system before deployment. Here we would need to specify:


After completing the ML task, Feature engineering, and the evaluation details, the next is to specify:

Such information is required to later decide on how to deploy the ML model.

Making Predictions

This block includes information about when we make a prediction on new inputs.

Collecting Data

Related to the Making Predictions, the Collecting Data block gathers information about new data that should be collected in order to re-train the ML model. In this way, we specify how we prevent the ML model decay phenomenon. Further questions to answer in this block are:

Building Models

Tightly related to the previous block, the Building Models answers questions regarding updating the ML models, because different ML tasks require different frequencies of model re-training:

Live Evaluation and Monitoring

After deployment, the ML model should be evaluated and here we would need to specify both model and business metrics, which should correlate. Generally, the metrics should follow the S.M.A.R.T methodology and be: Specific, Measurable, Achievable, Relevant, and Time-bound.

The deliverable in this stage is the completed Machine Learning Canvas. The effort to fill out this canvas might initiate an existential discussion regarding the real objective and hidden costs for the ML-software. Such a discussion might result in a decision not to implement AI/ML at all. Possible reasons can be as follows:

Another question would be when to deploy ML/AI? The following Figure shows the trade-off of early vs. late ML model deployment.

When to Deploy AI?

Further reading