AI in the Early Stages of Multifamily Projects
Updated: Aug 12
By Chloe Balkcom | 2026 Intern Research Program
Shapiro & Company Architects
Research Question
How can AI and data tools support feasibility, concept generation, and design intelligence in practice?
As the development of new technology arises, the AI world and the architectural world are colliding. From visualization and rendering to site planning to performance analysis, AI has become a powerful tool that can amplify design excellence, improve sustainability, and streamline project delivery.
This research explores and evaluates the capabilities of current AI and data tools in feasibility studies and site planning, to see how these tools can optimize Shapiro’s current workflow.
At Shapiro & Co., during this early-stage planning, we focus on feasibility studies, site planning, and concept generation. Beyond understanding their workflow, it was critical to gather what aspects of this process we would want AI to supplement. Through interviews with our principals and department heads, they described similar components shown in Figure 1.
Figure 1: Shapiro & Company AI Wants
From this chart, I created a matrix to easily evaluate the AI softwares, which I'll be referring to as the "Ideals Matrix."
Figure 2: Ideals Matrix
Throughout my exploration into current AIs, there were two standouts: Testfit and Autodesk Forma. They both are site planning platforms with AI capabilities. Testfit focuses more on being a real estate feasibility platform while Forma focuses on site and environmental analysis.
Testfit and Forma's Expectations Assessment
Testfit and Forma were evaluated against the Ideals Matrix (Figure 2) to determine where each platform exceeded, met, or fell short of company requirements.
Tesfit excelled in its Asset (user-made parameter) library, unit modification abilities, and file exports to Autocad. Testfit met expectations in site preparation and environmental data, zoning integration, AI assisted generative design, yield estimation, and platform organization and comparison. However, Testfit failed when met with excessive constraints during AI generation and when exporting the model to Revit.
Forma exceeded expectations with its parking generation tool, unit editing abilities, reusable unit libraries, and proposal site metrics. Forma was adequate in its Autocad file imports and site data library and integration. Forma was lackluster in its site production tools like boundaries and setbacks, Autocad file size limits, and Revit export feature.
Testfit & Forma's AI Breakdown
Testfit Scenario I:
Testfit Scenario II:
Testfit’s AI has no design eye; It requires a designer expertise to manipulate the AI’s solutions properly. The AI has multiple moments to impact the site planning process through regions, not just the initial solve, but the most integration occurs on the initial solve. The AI uses generation templates for the parameters to make proposals. These templates include many of the necessary parameters for proper site planning such as unit mix, heights, parking requirements, and more. It chooses solutions based on how well it meets criteria not by proper design principles. The AI sorts the solutions by criteria and the number of options 10, 50, to infinity based on the user’s choice. It can produce solutions that do not follow all of the parameters exactly, but it is explicit on when it fails.
Forma Scenario:
Forma doesn’t have an AI. It computes its automatized generation through this process. There are 3 layout organizations, 4 building types, and 3-4 options generated from a combo of layout and building type. This totals to 36-48 generated options total. Forma’s automated solutions are bad design solutions. There is a lack of building typologies; it requires the user to manipulate the buildings to create the layout they want. The parking tool is good. It follows parameters and easily manipulates in a similar way to Testfit. Most of the site plan’s elements were placed manually and were completed post-automation. The software’s tool made element application smoother by using user-set parameters to manually apply. These components are ones Shapiro wanted to maintain full autonomy over, not to supplement with AI.
Testfit & Forma's Evaluation
Testfit
TestFit is weaker in criteria 1-3 than initial thought. This is due to some kinks in the software, and the unknown of how powerful the AI engine in the software is. It is not able to discern like typical engines such as Gemini, ChatGPT, etc. There aren't options to give feedback on “good” and “bad” designs, like other engines. It is not a smooth, intuitive process and requires a learning curve to understand. It is an intense platform to pick up and will require dedication and quick-learning to apply effectively, but once done, it can be a plug and play. The functions behaved differently than advertised. In reality, they are clunky and have a tendency to fail. The asset library is critical in making TestFit a well-rounded platform. It permits easy replication through different schemes and evaluations/projects. Its integration capabilities with AutoCad and Revit was a mixed-bag of good and bad. It was positive with exports to Autocad and imports from Revit, but it was lackluster in exports to Revit and imports from Autocad. Overall, Testfit met my expectations with slight differences than anticipated. It is a dense, complex tool but with finding a solid workflow it can be used for quick feasibility studies.
Forma
Forma was less of a well-rounded platform than anticipated. There was no AI, and it did not generate good designs by Shapiro standards. There is potential with storage functions such as the local and central libraries. It had disappointing Revit integration due to the new nature of the Revit plug-in as a BETA feature. When this gets solved, it can be a great starter tool. There are significantly less requirements and parameters available to use, but it follows one’s given. It has lackluster site planning tools i.e. boundaries and setbacks for a site design platform. This is not an AI platform; it is a 3D visualization tool with automation that speeds up the process and metrics that can save time.
Recommendations to Apply to the Current Process
It is important to see if working within tools like Forma and Testfit will be optimal for the current workflow. I do not see either of these platforms as necessary. The AIs in these fields are not advanced enough to be completely altering to current workflows, but they would save time. If there are hesitations to current workflows, I’d opt to explore the 3D capabilities and offerings of Forma, and then, make the jump to TestFit when a generic workflow is established to easily translate across platforms. TestFit could be a great platform for Shapiro, but it would depend on the effort made and the time dedicated into preparing the tool to be as efficient as it could be. With how the AI market is currently, the perfect tool could be around the corner while restrictions and regulations are around the corner too.
Sources:
