Gordon Ritter: I predicted AI’s learning loop a decade ago. The doomers are still measuring the wrong thing

The article argues that enterprise AI will create durable competitive advantage by capturing internal 'traces' of how employees think, decide, and work, rather than merely automating tasks. It cites examples from McKinsey, Bain, EY, Ramp, and Veeva to support the case that AI is best used inward to compound proprietary knowledge. The piece also highlights data- and process-leakage risks from public AI tools, including Anthropic’s changed consumer terms and a reported 77% of employees pasting data into generative AI tools.

Analysis

The market is still treating enterprise AI like a generic productivity wave, but the more important effect is rent extraction from workflow ownership. The winners are firms that can instrument human decision-making inside proprietary systems and convert every correction into a training loop; that creates a data moat that is both operational and contractual. That is structurally better than selling standalone models, because the value accrues where the work happens and becomes harder to replicate as the interaction history compounds.

RAMP is the cleanest public-market expression of this thesis. Its upside is not just seat expansion; it is the optionality from becoming the system of record for agentic workflows, where each new automated task improves retention and raises switching costs. VEEV has a similar profile in a more regulated vertical: if it becomes the layer where life-sciences decisions are made and logged, AI shifts from a feature to an embedded control point across the customer workflow, extending ARPU and reducing churn.

TRI and RELX face a slower but more dangerous threat: not near-term revenue collapse, but margin pressure from AI-native alternatives that can compress high-value research and workflow subscriptions into lower-cost bundles. The first-order selloff is probably too emotional, but the second-order risk is that public AI tools train customers to expect cheaper, semi-custom answers, which weakens pricing power over 12-24 months. That makes these names vulnerable to multiple compression even if reported fundamentals hold up initially.

The contrarian view is that the bear case is less about job loss and more about leakage of process IP; that risk is underpriced because it is invisible in traditional data-security budgets. The next catalyst is not macro adoption, but policy and procurement changes: once large enterprises formalize restrictions on public-model usage and push internal trace capture, the gap between AI leaders and laggards should widen quickly over the next 2-4 quarters.

Key Decisions for Investors

• Long RAMP vs. short a basket of legacy workflow software over 6-12 months; thesis is that agentic trace capture expands lifetime value and raises switching costs faster than consensus models reflect. Add on weakness if growth decelerates less than peers.
• Initiate a tactical long VEEV position into any post-earnings pullback, targeting a 3-6 month horizon; asymmetric upside comes from AI becoming embedded in regulated workflows, while downside is limited by recurring revenue durability.
• Reduce/short TRI and RELX on strength over the next 1-2 quarters; this is a multiple-compression trade, not a collapse thesis. Use call spreads or a pairs trade versus higher-quality software to limit fundamental bleed risk.
• Pair long RAMP / short TRI as the clearest expression of 'workflow moat' vs. 'AI-disintermediation risk' over 6 months; expect relative performance to track internal AI adoption and product velocity.