How do you diffuse AI into the real world? - Varun Shenoy, Long Lake

AI deployment in real-world services businesses requires a systematic progression from co-pilots to co-workers, leveraging enterprise data through learning l...

By Sean Weldon

Systematic AI Deployment in Services Businesses: From Co-Pilots to Co-Workers Through Enterprise Learning Loops

Abstract

The deployment of artificial intelligence in real-world services businesses represents a critical challenge in technology diffusion, comparable to the generational adoption timeline of electricity in the 1880s. This synthesis examines Long Lake's operational model for AI integration across 35 acquired services businesses, analyzing the systematic progression from co-pilot tools to autonomous AI co-workers. The research identifies three fundamental requirements for successful AI deployment: (1) a structured capability spectrum from simple retrieval-augmented generation to long-running autonomous agents, (2) enterprise-specific learning loops that leverage proprietary data and real-world task traces, and (3) in-person software-service co-design that transcends traditional vendor-customer relationships. Findings indicate that frontier models trained on public internet data cannot adequately handle the exceptions and tacit knowledge inherent in enterprise services work, necessitating continuous post-training on out-of-distribution business data and physical presence for enablement success.

1. Introduction

Despite remarkable demonstrations of artificial intelligence capabilities - from automated flight booking to code completion and customer service - real-world deployment in services businesses remains fundamentally unchanged. This gap between technological capability and practical adoption constitutes what may be termed the AI diffusion problem, which historical precedent suggests requires generational timescales to resolve. The adoption of general-purpose technologies such as electricity in the 1880s required not merely technological capability but comprehensive systemic transformation: infrastructure replacement, equipment acquisition, and workforce retraining.

Long Lake presents an alternative deployment model that addresses this diffusion challenge through direct ownership and operation of services businesses rather than traditional software vendor relationships. Having raised over $3 billion from investors including Elad Gil, General Catalyst, and AlphaWave, the organization has acquired 35 businesses across homeowner associations, property management, architecture, and human resources services, culminating in a $6.3 billion take-private of American Express Global Business Travel. The organization operates under an operator-owner model where AI performance failures become internal accountability issues rather than external customer complaints.

This synthesis examines Long Lake's systematic approach to AI deployment, analyzing the progression from assistive tools to autonomous agents, the mechanisms for leveraging enterprise-specific data, and the critical role of physical co-design in enabling adoption. The central thesis posits that successful AI deployment requires three interdependent components: a structured capability progression that builds trust through incremental autonomy, data flywheel mechanisms that capture real-world task complexity, and in-person co-design that addresses organizational resistance to process change.

2. Background and Related Work

2.1 The Jagged Frontier Framework

The concept of the jagged frontier describes the uneven distribution of AI capabilities across different task domains. While frontier models demonstrate remarkable proficiency in certain areas, performance degrades sharply in adjacent domains. This phenomenon is particularly evident when comparing code generation tasks to services work. In software engineering contexts, synchronous agents with file system access can effectively assist with desktop coding, while sandbox-wrapped asynchronous agents can autonomously generate pull requests. Engineers naturally parallelize work, comfortably managing ten concurrent asynchronous jobs.

In contrast, services work has traditionally been conducted serially - processing one email at a time rather than ten simultaneously. This fundamental difference in work structure presents both a challenge and an opportunity: how to represent knowledge work as code to leverage models' extensive code training, and how to develop asynchronous mechanisms for traditionally serial work. The frontier remains jagged precisely because most valuable enterprise tasks exist not in public internet content but in private enterprise data, institutional knowledge, legacy software, and tacit understanding.

2.2 Technology Diffusion Theory

The diffusion of general-purpose technologies historically requires more than demonstrated capability. Technology adoption necessitates systemic transformation including infrastructure replacement, capital equipment acquisition, and workforce retraining. This contextualizes AI deployment as fundamentally a diffusion problem rather than a capability problem, suggesting that the critical question shifts from "what can models do?" to "how do models complete economically relevant tasks in production environments?"

3. Core Analysis

3.1 The Co-Pilot to Co-Worker Progression

Long Lake's deployment strategy follows a structured capability spectrum that progresses through five distinct stages. The co-pilot stage implements simple retrieval-augmented generation chatbots providing quick information retrieval with basic system integration. This foundational layer establishes user familiarity without requiring process change.

The synchronous agent stage introduces real-time two-way interaction, exemplified by tools like Claude Code and Codex, which operate on 1-5 minute timescales and require explicit user triggering. The asynchronous agent stage represents a critical transition point, operating in background processes that can be triggered by external job queues rather than user queries. These agents proactively offer advice without user initiation, fundamentally changing the interaction model from reactive to proactive.

The long-running agent stage, currently identified as a core problem for research labs, involves agents operating across hours, days, weeks, or months. Finally, the AI co-worker stage represents a proactive partner completing work alongside users. The critical insight from Long Lake's experience is that organizations must "earn the right" to deploy more autonomous agents through successful implementation of earlier stages. Attempting to deploy co-worker-level agents without establishing trust through incremental progression results in adoption failure.

3.2 Enterprise Data Flywheel Mechanisms

Frontier models trained exclusively on public internet content face fundamental limitations when deployed on enterprise tasks. The most valuable work exists in proprietary data, institutional knowledge, legacy systems, and tacit understanding inaccessible during pre-training. Long Lake addresses this through a real-world data flywheel with four interconnected components.

First, agents collaborate with employees on actual work, generating rich traces that capture tool calls, failures, and edge cases. Second, these traces enable construction of real-world evaluations with ground truth data. Third, continuous hill climbing against these benchmarks produces iteratively better agents. Fourth, improved agents drive increased impact and adoption, generating additional traces. This flywheel structure creates a self-reinforcing cycle where each component strengthens the others.

The rich traces enable three critical capabilities. Automatically generated and scored evaluations incorporate both implicit feedback (derived from differences between AI-generated data and ultimately submitted data) and explicit feedback, eliminating the need for manual annotation. Post-training on out-of-distribution enterprise data enables task completion that frontier models cannot perform. Customization occurs at three levels: per-company (different processes), per-user (different work styles), and per-client (different service standards).

Critically, hill climbing benchmarks become regression tests on weekly cycles, creating continuous improvement mechanisms. This addresses the fundamental challenge that "the exceptions are the job" - real work is characterized by messy edge cases that constitute the actual task rather than obstacles to it.

3.3 The Continual Learning and Enablement Loop

Long Lake identifies two converging trends in AI deployment: continual learning (agent improvement via feedback) and enablement (enterprise adoption). Traditionally, these functions are organizationally siloed, with continual learning owned by research and platform teams while enablement is owned by growth and deployment teams. However, these represent the same underlying loop: agents improve only if used, and people use agents only if adoption is worthwhile.

This creates a snowball effect where initial usage is the critical blocker. Providing the best AI tool does not guarantee adoption if employees continue existing processes. Organizations with century-long operational histories demonstrate significant resistance to process change. An employee who has closed books using a particular method for 20 years will not spontaneously adopt new tools without intervention.

The synthesis of continual learning and enablement reveals that improvement and adoption are not sequential but simultaneous requirements. Technical capability improvements without enablement investment yield no deployment gains, while enablement efforts without continuous improvement fail to sustain adoption.

3.4 Extreme Software-Service Co-Design

Drawing inspiration from hardware-software co-design principles, Long Lake implements extreme software-service co-design possible only within unified organizational structures. This co-design operates at two levels: metaphorical and physical.

Metaphorical co-design involves embedding products natively into existing tools - Excel, ERP systems, 3D design software, Outlook, Gmail - to reduce enablement friction. By integrating into established workflows rather than requiring workflow replacement, adoption barriers decrease substantially.

Physical co-design requires in-person engagement that cannot be replicated through remote communication channels. Long Lake's deployment team, representing approximately 40% of the technology organization, conducts lunch-and-learn sessions, conferences, one-on-one training, and informal conversations during activities like mountain biking. The critical insight is that "you cannot co-design software with the services business over Zoom or over a support ticket. You have to be there. You have to be in person."

This physical presence enables what is termed "touching grass" - understanding real work as it exists outside technology hubs and conference rooms. The 100-year-old firms within Long Lake's portfolio operate with deeply embedded processes that require intimate understanding to successfully modify. Remote vendor relationships lack the contextual understanding and trust necessary for process transformation.

4. Technical Insights

The technical architecture for enterprise AI deployment reveals several implementation-critical insights. Asynchronous agents require external triggering mechanisms beyond user queries, specifically job queue integrations that enable proactive assistance. For code generation tasks, sandbox-wrapped agents can autonomously build, test, and generate pull requests, representing a largely solved technical problem.

Post-training on enterprise data that is out-of-distribution for frontier labs enables task completion impossible with base models. This necessitates infrastructure for continuous model fine-tuning on proprietary data sources. The collection of rich traces from agent-employee collaboration must capture not only successful completions but failures and edge cases, as these provide the most valuable training signal.

Implicit feedback mechanisms derive signal from differences between AI-generated outputs and final submitted data, eliminating annotation overhead. This requires careful instrumentation to capture both intermediate and final states. Weekly hill climbing benchmarks require automated evaluation infrastructure that can execute at scale, with successful benchmarks automatically converting to regression tests.

Customization at three levels - company, user, and client - requires hierarchical model architectures or parameter-efficient fine-tuning approaches that can maintain separate adaptations without prohibitive computational costs. The technical challenge involves balancing customization depth with inference latency and serving costs.

The representation of knowledge work as code leverages models' extensive code training, but requires careful abstraction design. Tasks must be decomposed into code-representable components without losing essential context. The development of asynchronous forking mechanisms for services work represents an open research problem, as the sandbox approaches successful in code generation do not directly transfer to customer service, financial processing, or administrative tasks.

5. Discussion

The findings from Long Lake's deployment model reveal fundamental tensions in AI adoption that extend beyond technical capability. The operator-owner model creates accountability structures impossible in vendor-customer relationships, suggesting that organizational structure may be as important as technical architecture for successful deployment. This raises questions about whether successful AI integration in services businesses requires ownership consolidation or whether alternative accountability mechanisms can replicate these incentives.

The convergence of continual learning and enablement into a single loop challenges traditional organizational structures in both AI companies and enterprises. Research organizations typically optimize for capability improvements measured on public benchmarks, while deployment teams optimize for adoption metrics. The recognition that these are interdependent rather than sequential suggests that organizational restructuring may be prerequisite to successful deployment.

The requirement for physical co-design presents scalability challenges. If in-person presence is necessary for enablement success, deployment velocity is constrained by the availability of skilled personnel who can operate effectively in both technical and operational contexts. This suggests a potential bottleneck in AI diffusion that cannot be resolved through improved model capabilities alone.

The emphasis on handling exceptions as the core job rather than obstacles reveals limitations in current evaluation methodologies. Public benchmarks typically measure performance on well-specified tasks with clear success criteria, while real-world services work is characterized by ambiguous requirements, conflicting constraints, and tacit knowledge. This gap between benchmark performance and production utility suggests that current frontier model development may be optimizing for the wrong objectives.

6. Conclusion

This analysis demonstrates that successful AI deployment in services businesses requires systematic progression through capability stages, enterprise-specific learning loops, and in-person co-design rather than mere access to frontier models. The operator-owner model creates accountability structures that drive genuine problem-solving rather than capability demonstrations. The co-pilot to co-worker spectrum provides a structured path for building organizational trust and technical capability simultaneously.

The real-world data flywheel addresses fundamental limitations of models trained exclusively on public internet content, while the synthesis of continual learning and enablement reveals that improvement and adoption are inseparable requirements. Physical co-design emerges as non-negotiable for overcoming organizational resistance in established services businesses.

For AI researchers and engineers, these findings suggest that deployment success requires capabilities beyond model performance: infrastructure for continuous post-training on proprietary data, mechanisms for capturing and learning from rich task traces, and organizational structures that unite technical development with operational deployment. For services businesses, the implication is that AI adoption may require ownership models rather than vendor relationships, and that enablement investment is as critical as capability acquisition. The AI diffusion problem will be resolved not through better models alone, but through systematic integration of technical capability, organizational structure, and physical presence.

Sources

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub