Deploying AI Agents into Real-World Services Businesses

The Strategy of Vertical Integration

Long Lake approaches AI deployment by acquiring services businesses—spanning property management, architecture, and HR—to act as both the operator and the owner of the technology. This structure eliminates the vendor-client blame game, forcing the engineering team to build agents that solve actual operational problems rather than just demoing capabilities. The core thesis is that AI diffusion follows the same slow, labor-intensive path as electricity in the 1920s: it requires ripping out legacy systems, retraining staff, and physically co-designing software with the people performing the work.

Moving from Copilots to Coworkers

The transition from simple RAG-based chatbots to autonomous agents requires a ladder of autonomy. While synchronous agents (like coding assistants) are well-understood, the frontier lies in asynchronous agents that can handle long-running tasks across days or weeks. Because services work is traditionally serial, the team focuses on representing knowledge work as code to enable parallelization. This involves building custom integrations with legacy ERP systems and 3D design software, rather than relying on generic browser-based interfaces.

Closing the Loop with Ground Truth

Because the most valuable tasks—such as scoping construction blueprints or coordinating vendor repairs—are not documented on the internet, the team generates proprietary data through internal operations. Every interaction between an employee and an agent creates a rich trace of tool calls, errors, and corrections. This data serves two purposes: it creates objective ground truth for automated evals (e.g., did the roof get repaired?) and provides a dataset for post-training models on out-of-distribution tasks. This creates a flywheel where usage drives continual learning, which in turn improves the agent and encourages further adoption.

The Reality of Adoption

Technology adoption in century-old firms fails if it relies on remote support or Zoom calls. The team emphasizes "touching grass" by embedding engineers directly into the business units, attending trade conferences, and building tools that live inside the software employees already use, such as Excel or Outlook. The goal is to lower the activation energy for non-technical users by making the AI a proactive partner that understands the specific, messy exceptions of the business.