THE MOVING FRONTIER · MAY 2025 RETROSPECTIVE
Data is not a strategy
For years, software businesses were told that their data would become their moat. There is truth in that claim, but it encourages lazy thinking. Access to records is not the same as knowing which decisions matter, what good looks like, or how a product should improve.
In AI products, the stronger advantage may be a learning loop: a customer attempts meaningful work, the system produces an outcome, an expert provides a signal about its quality, and the product becomes more useful for the next attempt.
That loop is difficult to copy. It depends on workflow position, earned trust, and the discipline to translate feedback into product change.
Four requirements for compounding
First is workflow position. The product must participate in a task customers repeatedly perform, not sit on the edge as an optional assistant.
Second is a quality signal. A thumbs-up is weak evidence. Did the user accept the recommendation, correct the classification, redo the work elsewhere, or escalate because evidence was missing? Signals should be tied to real outcomes.
Third is a mechanism for improvement. Better retrieval, clearer instructions, tooling, evaluation cases, or interaction design might all be candidates. ‘The AI will learn’ is not an improvement plan.
Fourth is permission. Sensitive customer information should not quietly become a training asset. Privacy boundaries, explicit rights, and appropriate controls determine which signals can legitimately be reused.
The boring details become strategic
Consider a system that helps assemble compliance evidence. The insight is not simply that someone uploaded a document. It is which evidence a reviewer accepted, why an item failed, which citation resolved a question, and what exception required human judgment.
None of these signals is glamorous. Together, they describe the shape of useful work. Capturing them appropriately can improve the surrounding workflow and the next generation of evaluations.
The compounding comes from product design choices: what gets logged, how corrections are represented, whether failure modes become tests, and whether the system can learn without putting the customer at risk.
Beware the wrong loop
A learning loop can optimize a bad objective. Reward quick approvals and a system may produce work that looks complete but is poorly supported. Learn only from power users and the product may become unusable for everyone else.
A trustworthy loop periodically asks whether the output held up under independent review. Domain experts, careful sampling, and an explicit definition of success matter more than feedback volume.
Models and interfaces will keep changing. The durable advantage is not data sitting still. It is the organization’s ability to learn responsibly from the work.
Editorial note: This is a retrospective essay written in November 2026. May 2025 describes the period explored, not its original publication date.
