Overview
AI models can already perform tasks that looked like science fiction two years ago, yet many real businesses remain largely unchanged. Varun argues that this gap is normal: like electricity, a general-purpose technology produces broad economic value only after organizations replace equipment, redesign processes, and retrain people. Long Lake addresses this diffusion problem by acquiring and operating service businesses rather than selling them software from the outside, making failed deployments its own operational problem. Its approach has three connected parts. First, agents should progress deliberately from informational co-pilots to synchronous, asynchronous, long-running, and eventually proactive co-workers; greater autonomy must be earned through capability and trust. Second, collaboration with employees produces traces, feedback, outcomes, and evaluation data unavailable on the public internet, enabling better agents and internally post-trained models. Third, adoption and continual learning must operate as one loop: usage creates learning data, improvements make the agent more useful, and usefulness encourages more usage. The initial adoption problem remains decisive. Long Lake’s proposed answer is extreme software-service co-design—embedding tools in existing systems, working directly with employees, and learning in person how exception-filled service work actually happens.
Sections
Strategic Insights
Higher-order implications derived from Long Lake’s operating model and deployment experience.
- The scarce resource in enterprise AI may be access to authentic workflows and outcome data rather than access to frontier models. Ownership gives Long Lake both the operational visibility and the incentive structure needed to convert tacit work into evaluation and training assets.
- Agent autonomy is partly a social permission system. Even when technical capability exists, organizations must accumulate evidence, trust, and workflow familiarity before employees will accept more independent execution.
- Service businesses may need industry-specific forms of asynchronous work rather than a universal agent interface copied from software engineering. The correct trigger, workspace, review mechanism, and unit of parallelization will depend on the underlying profession.
- The long tail of exceptions can become a compounding advantage: every observed correction or edge case can be converted into an evaluation, regression test, customization rule, or post-training example.
Key Comparisons
Contrasts used to explain why enterprise deployment differs from an AI demonstration.
- AI demonstrations reveal isolated capability, whereas operational deployments must survive legacy systems, human habits, exceptions, and accountability for real outcomes.
- External vendors can attribute failure to customer implementation, while operator-owners remain directly responsible when an agent does not work.
- Coding work already supports sandboxed, parallel asynchronous agents, while service work is traditionally serial, context-heavy, and difficult to isolate.
- Continual learning improves the agent, while enablement creates the usage that supplies learning data. Treating them as separate functions breaks their natural feedback relationship.
Deployment Lessons
Practical principles drawn from operating AI inside acquired service businesses.
- Increase agent autonomy incrementally and only after both performance and user trust justify the next stage.
- Capture complete work traces, including failures, corrections, tool calls, and final outcomes, because they are the raw material for evaluations and improvement.
- Judge agents against completed business outcomes rather than response quality alone.
- Embed AI into existing systems and workflows to reduce the effort required for employees to begin using it.
- Work directly with employees in their operating environment because tacit practices and workflow exceptions are difficult to discover remotely.
Forecasts
Forward-looking claims made or strongly implied by the speaker.
- AI diffusion will remain one of the most important technology and organizational problems over the next two decades.
- Agents will progress from synchronous and asynchronous tools toward long-running systems and ultimately proactive AI co-workers.
- Asynchronous agents for service industries will become a major development frontier, but their form factors will differ substantially across professions.
- Enterprises that unite adoption with continual learning will create compounding improvement loops in which greater usage produces better agents and better agents produce greater usage.