Principle AI Product Building Business Models Future of AI

Proprietary feedback loops create moats that widen with every interaction

When usage generates data that competitors cannot replicate — correction patterns, preference signals, domain-specific edge cases — the product improves faster than any new entrant can catch up

Nikunj Kothari — Revealed Preferences · Mar 16, 2026 · 16 connections

The strongest AI moat isn’t proprietary data at rest — it’s proprietary data in motion. Every user interaction generates correction patterns, preference signals, and edge-case resolutions that feed back into the system. A competitor starting from scratch faces not just a data gap but an accumulating gap: the incumbent’s system improves with every interaction while the challenger has zero signal.

This goes beyond Domain-specific skill libraries are the real agent moat, not core infrastructure (which captures static expertise) to dynamic expertise that self-updates. Skills encode what you knew yesterday; feedback loops encode what you’re learning today. The mechanism works through Compound engineering makes each unit of work improve all future work applied to the product itself — each user session doesn’t just complete a task but improves future task completion. And Cross-user knowledge transfer works without fine-tuning — just a database and prompt engineering shows the infrastructure is surprisingly simple: a database and prompt engineering, not expensive training pipelines. The key is designing the capture mechanism from day one — retrofitting feedback loops onto a product that wasn’t built for them is orders of magnitude harder. This is fundamentally a UX problem, which is why The data flywheel is a UX problem — only vertical workflow surfaces can capture the knowledge: the flywheel only spins if the workflow surface is shaped to capture the right signal, and horizontal tools structurally can’t shape it. The capture mechanism also has to live somewhere you control: Closed harnesses behind APIs create memory lock-in by design argues that the feedback loop itself is the memory layer, so ceding that layer to a closed harness surrenders the moat. Nadella elevates this loop to the firm’s defining asset in You can offload a task, or even a job, but you can never offload your learning — a hill-climbing machine where every improved workflow generates better training signal, compounding into an advantage that survives any new model.

Connected Insights

References (6)

→ Compound engineering makes each unit of work improve all future work → Cross-user knowledge transfer works without fine-tuning — just a database and prompt engineering → Domain-specific skill libraries are the real agent moat, not core infrastructure → Closed harnesses behind APIs create memory lock-in by design → You can offload a task, or even a job, but you can never offload your learning → The data flywheel is a UX problem — only vertical workflow surfaces can capture the knowledge

Referenced by (10)

← You can offload a task, or even a job, but you can never offload your learning ← The data flywheel is a UX problem — only vertical workflow surfaces can capture the knowledge ← Agent edits are automatic decision instrumentation — every human correction is a structured signal ← Compound engineering makes each unit of work improve all future work ← Domain-specific skill libraries are the real agent moat, not core infrastructure ← The trace→eval→harness flywheel compounds agent quality — every production interaction generates its own training data ← The context flywheel is a Day 90 moat — Day 0 comparisons are misleading ← Closed harnesses behind APIs create memory lock-in by design ← When products and models converge, the moat is the institution — how a company attracts people, distributes authority and compounds judgment ← Point a second agent at the first agent's failure log and the system repairs itself overnight