How to Build an AI Blueprint
Developing artificial intelligence (AI) products is a hefty investment and high-risk activity. It is the best practice to conduct an AI Blueprint before developing an AI product to de-risk issues down the road. This approach allows organizations to move quickly and adjust accordingly. The early determination will help to commit and deliver what is planned. During an AI blueprint, the aim is to answer seven key questions. There are many other articles referenced here that discuss the approach in here. AI project canvas, AI Cavas, and Machine learning Canvas, Derisking AI by design are well-described techniques that we primarily use for this blog.
1) Objective and output
Organizations are often trying to define the solution approach versus understanding the objective and desired outcome. Understanding the solution’s aim, the business’ and users’ challenges help design the solution fitted to the users’ needs. Understanding the objective and expected output leads us to a better and more accurate blueprint of the development work. In the Prediction Machines book, AI canvas is one of the best techniques to build the AI task’s details. This includes objective, prediction, judgment, action, outcome, input, training, and feedback, which are key AI tasks, that need to be defined before developing any product.
2) Value and Performance
We need to measure and refine the estimated value and potential revenue, post-deployment of the product to measure return on investment (ROI). By analyzing the data, conducting research, and interviewing business experts, the estimated value is measured. We also need to study and define how we can measure the solution models performance and key performance metrics (KPIs). Businesses usually would like to improve their operations using advanced analytical solutions. Building a baseline to compare the new solution and measure the solutions improvements, help organizations measure their actual value and desired performance. These items are used to plan for offline and online evaluation according to Machine learning Canvas.
3) Data Assessment
Data is the foundation of solving an analytical problem and building an AI product. Before we develop an AI solution, we conduct a data assessment. Data assessment includes data availability, data quality, and data sufficiency. We need to list the data required to solve the problem and analyze if we have the available data. If the data is unavailable, the plan is to purchase or build the data. We also need to run tests by looking at the data quality and conduct interviews with data experts pertaining the specific industry. The models often require some amount of historical data to predict accurately.
4) Analytical Approach
We need to understand the high-level problem solution for the proposed challenge. Companies often predetermine the solution without understanding the problem and its associated challenges. Some problems can be solved by digital and not analytical solutions. The problem modeling and solution design are critical phases to design an analytical approach to solving the problem. We design different models and model the relationship between the models. This process is called model mapping. We also need to define the proposed solution alongside each model’s complexity, input, and output. The solution’s design will help map out the AI blueprint and AI development more effectively.
We need to study and analyze the data to identify the predictability or optimality of the solution. Some problems may be difficult to predict since there is not any data-driven and eventful pattern. This issue can happen based on some random events that are hard to predict. Looking at the data set, analyzing the trend, fluctuation, and noises of the data, we can build a simple machine learning model. The machine learning model can help us understand how to solve the problem and define expected value gain much better.
6) Business and Architecture Integration
Integrating an AI solution in the existing business unit and architecture system is challenging. It is essential to identify how the existing solution will integrate with the existing systems and current infrastructure. We need to consult with the organization’s enterprise architectures to integrate complexity, identify the risks, and any pre-existing activity before initiating AI product development. We also need to identify and understand how the solution will change the business operations and impact its workflow. Early understanding of these elements will help organizations address the risks and complexity of change management and user adoption. These
7) Workplan, Resources, and Timeline
Finally, we need to complete AI blueprints that define the work items alongside required skills and estimated timeline and cost. The work items are defined based on Intelius AI practices and studies from data assessment, analytics approach, and integration analysis. We also often propose prework activities before we launch AI development and ensure the entire team is engaged. Once we have defined work items and pre-activity activities, we build the proposed resources and skills for the development phase and execute it. The next item is to identify the timeline, including significant milestones and deliverables. It is a requirement that we pinpoint the timeline and budget constraints and expectations for AI implementation from the business and user perspective to outline the implementation targets.
All these items are required to make sure we can kick off the AI development and guarantee success. Intelius helps you with free AI consultancy to support you describe our approach to make your transition much more successful.
Agarwal, Ajay, and Avi Goldfarb. “Prediction Machines”: The Simple Economics of Artificial Intelligence, edited by Joshua Gans, Harvard Business Review Press, 2018, pp. 74–75.
Jan Zawadzki, “Intorducing the AI Project Canvas” https://towardsdatascience.com/introducing-the-ai-project-canvas-e88e29eb7024
Louis Dorard, “The Machine Learing Canvas” https://www.louisdorard.com/machine-learning-canvas
Juan Aristi Baquero, Roger Burkhardt, Arvind Govindarajan, and Thomas Wallace, “Derisking AI by design: How to build risk management into AI development” https://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/derisking-ai-by-design-how-to-build-risk-management-into-ai-development#