This paper introduces the LeanAI method, a structured approach designed to help Architecture, Engineering, and Construction (AEC) practitioners effectively plan AI implementations and bridge the gap between high-level business goals and technical execution (Agrawal et al., 2023 ). The core problem addressed is the high failure rate (cited as 85%) of big data and AI projects in the AEC industry, largely attributed to a disconnect between those planning AI initiatives and those implementing them. Practitioners often confuse what AI should solve (business needs), what it can solve (technical possibilities), and what it will solve (practical capabilities based on available resources), leading to unrealistic expectations and poorly defined projects.
The traditional approach to AI planning often starts and ends with high-level business needs (e.g., reduce costs, speed up projects) without clearly defining how AI will achieve these goals or whether the necessary prerequisites (data, algorithms, expertise) are available. This lack of clarity hinders implementation teams.
The LeanAI method, inspired by the Last Planner System used in lean construction, provides a framework to explicitly define and align key components early in the planning phase:
- Business Need (Necessity - What AI SHOULD solve): This defines the high-level objective or necessity driving the AI project. It focuses on the "bigger picture" impact required to gain management support and resources. Examples include reducing project costs, enhancing safety, or improving maintenance efficiency. The paper stresses the need for specificity here (e.g., "reduce maintenance cost" is better than "improve project performance").
- AI Problem Statement (Possibility - What AI CAN solve): This translates the business need into specific, well-defined problems that AI algorithms are capable of addressing. Multiple potential problem statements might be formulated for a single business need. This requires input from both domain experts (understanding business operations) and AI experts (understanding AI capabilities). For instance, to reduce maintenance costs, possible AI problem statements include "detect cracks on highways using images" or "predict traffic growth."
- Available Data, Labels, and Algorithms (Capability - What AI WILL solve): This component assesses the practical feasibility of solving the defined AI problem statement based on the organization's resources. It involves evaluating the availability and quality of required data, necessary labels for supervised learning, and the technical expertise needed to develop and implement the chosen AI algorithms. This step often requires input from IT teams and AI developers to get realistic estimates. Lack of sufficient data or expertise might invalidate certain problem statements.
- Metrics (Accountability - What AI DID solve): This involves defining two types of metrics to measure success:
- AI Metric: Measures the technical performance of the AI model (e.g., accuracy, precision, recall, F1 score) on the specific AI problem statement.
- Business Metric: Measures the impact of the AI solution on the original business need (e.g., dollars saved, percentage reduction in rework, safety incidents avoided). Crucially, the method emphasizes establishing a clear link between these two metrics to understand how improvements in AI performance translate into tangible business value. Unrealistic expectations regarding AI metrics (e.g., demanding 99%+ accuracy when lower might suffice or be the only achievable target) can doom projects.
The paper proposes a five-step iterative process for applying the LeanAI method:
- Define the Business Need: Start with a specific and precise business objective.
- Formulate AI Problem Statements: Brainstorm multiple specific tasks AI could perform to address the need. Involve both business and AI stakeholders.
- Evaluate Feasibility (Data, Labels, Algorithms): Assess if the necessary resources exist to solve the formulated problems. Get realistic estimates from experts.
- Define Metrics: Set both AI and business metrics and establish their relationship. Ensure targets are achievable.
- Iterate: Continuously evaluate performance against metrics and refine the business need, problem statement, or approach as needed.
A case paper involving highway maintenance illustrates the method. Initially, practitioners struggled with vague goals like "reduce costs." Using LeanAI, they defined the business need ("reduce maintenance costs") and brainstormed AI problem statements ("detect cracks," "predict crack growth," "predict traffic growth"). Evaluating feasibility revealed insufficient data for "predict crack growth." Further analysis linking AI metrics (accuracy) to business metrics (cost savings) showed that even high accuracy in "predict traffic growth" wouldn't yield significant savings. This led them to focus on the "detect cracks" problem, which was both feasible with available resources and impactful for the business need.
The method was validated in a workshop with 30 AEC practitioners. Most found it very or somewhat useful, appreciating its simplicity, practicality, and emphasis on aligning the different components. The development itself was based on ethnographic-action research involving longitudinal case studies and direct observation of AI implementations in AEC firms.
In essence, LeanAI provides a practical planning framework that forces explicit consideration of business needs, AI possibilities, resource capabilities, and measurable outcomes, thereby increasing the likelihood of successful AI implementation in the AEC sector.