Why is it so hard to get ROI from AI? Because building from first principles isn’t easy

The article is a conference recap centered on AI adoption, with executives from State Street, Deloitte, Wells Fargo, Anthropic, Mistral, Snowflake, Hyatt, and others emphasizing that companies must do more foundational workflow and governance work before expecting ROI. Wells Fargo said AI tools helped drive a 25% increase in new account openings, but most discussion was qualitative and strategic rather than about immediate financial results. Overall, the piece is informative and industry-focused, with limited direct market-moving content.

Analysis

The market takeaway is not that AI is slowing; it is that the next leg of monetization shifts from model novelty to workflow ownership. That favors platforms with embedded distribution and governance, because enterprises are starting to buy control planes, auditability, and integration layers rather than raw inference horsepower. In practice, that is more durable for SNOW and STT than for vendors selling generic “AI capability,” while the biggest upside for WFC comes from monetizing productivity inside a highly regulated base rather than from external AI product revenue.

The second-order effect is that AI value creation will likely be lumpy and delayed, which compresses near-term expectations for horizontal software names but extends the runway for infrastructure and workflow winners. If companies are re-architecting processes from scratch, the spend pool shifts toward data plumbing, permissions, identity, model routing, and oversight tooling before it shifts into broad application rollouts. That means the first beneficiaries are the picks-and-shovels, not the agents themselves, and vendors that can prove governance plus measurable ROI should take share from point solutions and services firms.

A key contrarian point: the current skepticism around enterprise AI ROI is probably directionally right in the next 2-4 quarters, but wrong on a 2-3 year horizon. The gap between experimental pilots and scaled deployment is exactly where valuation dispersion should widen; names tied to measurable workflow conversion can rerate while overhyped “AI-native” narratives fade. For banks and asset managers, the value of AI may show up more in loss avoidance, employee leverage, and customer retention than in immediate revenue, which means consensus will likely understate the compounding effect until it is visible in operating margins and cross-sell data.

Key Decisions for Investors

• Long SNOW vs. short a basket of unprofitable AI application names over the next 6-12 months: governance, data unification, and workload orchestration should outperform as enterprises defer broad agent deployment; target 15-20% relative outperformance if IT budgets stay disciplined.
• Add to WFC on pullbacks for a 6-9 month horizon: the market is likely underestimating AI-driven operating leverage in regulated workflows; risk/reward improves if management can translate productivity into margin expansion rather than just growth anecdotes.
• Buy STT as a medium-term beneficiary of AI control-plane demand: asset owners and regulated clients will pay for data lineage, permissions, and model governance, creating a higher-quality, stickier revenue stream than generic AI exposure.
• Avoid chasing software vendors that are positioning as pure AI wrappers unless they can prove workflow redesign and measurable unit economics within two quarters; if they cannot, expect multiple compression as ROI scrutiny intensifies.
• Consider a pairs trade long SNOW / short high-multiple SaaS names exposed to AI disappointment; if enterprise adoption remains cautious, infrastructure spend should prove more durable than application-layer monetization.