Software developer Robert Glaser warns that enterprise deployments of tools like ChatGPT and GitHub Copilot are creating isolated individual productivity gains while failing to build shared organizational knowledge. Despite companies committing budgets as high as €2 million on enterprise licenses, critical operational learnings remain trapped within individual work loops. Organizations must evolve beyond tracking token spend and adoption metrics to implement feedback harnesses that transform private AI breakthroughs into shared capabilities.
In a recent essay, Glaser references author Ethan Mollick’s framework from Making AI Work: Leadership, Lab, and Crowd, emphasizing that personal productivity increases do not automatically become corporate assets. While individual workers quietly become faster at analyzing data, writing code, or automating tasks, the broader organization learns almost nothing from these isolated breakthroughs.
The 'Messy Middle' of Enterprise AI Adoption
Many enterprises have entered what Glaser describes as the "messy middle" of AI implementation. In this phase, licenses for tools like GitHub Copilot, ChatGPT Enterprise, Claude Code, and Cursor are widely distributed, but usage patterns remain fragmented across teams.
While leadership tracks prompt counts or steering committee metrics, real execution varies drastically across departments:
- Tight feedback loops: Senior engineers use agentic tools to complete root-cause incident analyses in under an hour instead of two weeks.
- Loose delegation: Product managers prototype working software directly, bypassing traditional Figma screen mockups.
- Ad-hoc automation: Support teams independently convert recurring helpdesk tickets into automated workflows without coordination from central AI enablement teams.
Individual productivity gains from AI tools do not automatically translate into organizational learning. Instead of institutionalizing these breakthroughs, traditional corporate change machinery—such as monthly brown-bag demos, champion networks, and enablement decks—moves too slowly to capture fast-moving developer patterns.
Shifting From Token Metrics to Loop Intelligence
As enterprise AI usage grows, financial scrutiny will inevitably shift from open-budget experimentation to usage-contingent pricing and token budgets. However, Glaser cautions that measuring success purely through "token-to-output" metrics, such as counting generated pull requests, repeats old managerial mistakes.
Instead, companies need to evaluate "token-to-learning" outcomes to understand which decisions improved, which root causes were identified faster, and which product prototypes successfully eliminated weak ideas early.
The primary obstacle to this shift is legacy software management. Traditional Scrum framework ceremonies were designed for an era when human iteration was expensive and slow. Agentic workflows lower iteration costs, moving the primary operational constraint from raw code implementation toward intent, verification, and human judgment.