The cleanup trap: Stop asking RAG to fix bad data

The tradeable implication is not that AI demand is fading; it is that the spend mix is rotating toward the unsexy control layer. That should favor data observability, ingestion, lineage, identity, and policy enforcement vendors over names monetizing the model/app layer alone. In practice, the winners are likely to be the infrastructure and security names that get pulled into remediation projects after failed pilots, because once an enterprise discovers the pipeline is the bottleneck, the fix becomes recurring budget rather than a one-time experiment.

The second-order loser set is software companies that have priced AI as an attach-rate accelerator but do not control the underlying data plane. Their risk is a longer lag between pilot and production, which stretches sales cycles and delays expansion revenue by 1-2 quarters; that matters most for high-multiple enterprise software where the market is underwriting near-term AI monetization. The competitive effect is also negative for point solutions that rely on prompt-layer cleverness, since procurement teams will increasingly demand lineage, access control, and auditability before allowing production rollouts.

Contrarianly, this is not a blanket bearish call on enterprise AI capex. The consensus underestimates how much of the wallet share shifts from experimentation to infrastructure hardening, but it may be overestimating the speed of abandonment for AI programs. The main falsifier is evidence that enterprises are scaling AI into production without meaningful cleanup cycles: faster conversion rates, lower incident rates, and management commentary that governance spend is stabilizing rather than rising over the next 1-2 quarters.