How to evaluate and prioritise AI opportunities with the AI canvas

A guide to using the tools from the Prediction Machines book by Agrawal et al.

Identifying and harnessing AI opportunities is crucial for today’s digital innovation. The AI canvas is a powerful tool to evaluate the feasibility and potential impact of AI-driven solutions.

Use this article as a user-friendly quick start guide to systematically break down and analyse the potential of AI for your digital transformation use case.?

#aidesign #aicanvas #predictiveai #effectiveai #realworldai #augmentedintelligence

Getting started with the AI canvas

Download my adapted AI canvas template from https://shorturl.at/EN9tI

It is enriched with usage notes for impactful AI. And it adds two questions to frame and contextualise the core concepts.

The AI canvas is a lean template for creating a structured AI business case by dissecting the essential components of an expert workflow. It emphasises augmentation of existing workflows, what I like to call machine-in-the-loop human intelligence.

Read on for an explanation of the core concepts: Why use AI for your use case? How to build and maintain it? Once you’ve understood these, the short guide at the end steps you through applying the template to your problem.

Use AI when predictions lead to better decisions

The first objective is to characterise how machine prediction can improve decision making. Start by looking at the workflow as currently performed or as anticipated. Identify tasks and decisions within that workflow.

Key workflow concepts:

• Tasks: The fundamental unit of AI workflow transformation, tasks are collections of decisions. ?Our goal is to find decisions that can be automated or improved with better prediction.
• Decisions: Decisions are made based on judgment and may also incorporate predictions. Decisions may share common predictions but differ in the consequent actions.
• Jobs: Human roles or occupations. Experts employ prediction machines to help them make judgments about what action to take. Accuracy and efficiency gains lead to better outcomes.

Question: Should you take your umbrella?

To illustrate, let’s consider a decision-making scenario: whether to bring an umbrella when leaving the house. The actions are simple. You either take your umbrella or you leave your umbrella.?

Assume you don’t have a weather forecast but you do know that there’s a 25% chance of rain at this time of year. This is your prediction. We can use a decision tree diagram to visualise how decisions and predictions map to outcomes.

Answer: Why yes, you should take your umbrella!

Judgment refers to how you incorporate this information when making your decision. It is based on expected utility, accounting for both the likelihood of rain and your preferences for being dry and carrying an umbrella.

Let’s say you prefer being dry without an umbrella most, so assign the highest score to this outcome. Carrying an umbrella is ok. And you don’t like being wet. Given these settings, you take your umbrella because it has higher expected utility.

Take a moment to reflect what would happen if we had a different prediction or different preference scores. A higher likelihood of rain would reinforce the current decision. On the other hand, the decision would flip if you hated toting an umbrella.

AI requires training and feedback

The second objective is to characterise the requirements for operating, building and maintaining your predictive AI. These are the fundamental requirements for feasibility. At this point, it’s useful to look at the anatomy of a task.

Key data concepts:

• Input: Data required for an AI model to make a prediction.
• Training: Data required to train the AI, generally historical examples of input plus high-quality gold standard labels.
• Feedback: Continuous monitoring and improvement of the AI system to maintain and enhance prediction accuracy over time.

A simple 4-step guide for applying the AI canvas

Follow these steps to use the AI canvas:

Step 1: Identify tasks and decisions

Begin by analysing the workflow you aim to improve. Break it down into specific tasks and identify the decisions and actions within those tasks.

Example: In an MBA recruitment process, tasks might include soliciting applications, ranking candidates and making offers. Decisions involve determining who to interview or when to make an offer.

Step 2: Map predictions to judgments

Determine the key predictions required for each decision. Think how they will augment human judgment when deciding what action to take.

Example: Predicting the likelihood of a candidate accepting an offer can inform whether and when to make that offer.

Step 3: Define data requirements

Specify the data needed to train the AI model and generate predictions. This includes both input and training data.

Example: Historical data on past candidates and their outcomes can be used to train the AI.

Step 4: Incorporate feedback mechanisms

Establish how you will monitor and improve the AI model. Feedback is crucial to adapt to changes and maintain the model's accuracy.

Example: Continuously tracking the outcomes of decisions and updating the model with new data ensures it remains effective.

Go forth and digify with impact!

Download my adapted AI canvas template from https://shorturl.at/EN9tI

Note the canvas is necessary but not sufficient. It helps you decide whether you can build AI. To continue assessing whether you should, I recommend thinking carefully about your business plan as well as end-to-end design and validation.?

You are now equipped to identify and prioritise AI opportunities with a structured approach that ensures your AI solution is both impactful and feasible. Happy innovating!

What do you think?

Any suggestions for this guide? What tools do you use for AI strategy? How do you ensure solutions are safe, effective and human-centred? What should I write about next? Please share your thoughts in comments or DM. I appreciate feedback.

Acknowledgement

Thanks to Anna Pakarinen for expert content design advice.